Daily Papers

The adoption of Generative AI (GenAI) in various applications inevitably comes with expanding the attack surface, combining new security threats along with the traditional ones. Consequently, numerous research and industrial initiatives aim to mitigate these security threats in GenAI by developing metrics and designing defenses. However, while most of the GenAI security work focuses on universal threats (e.g. manipulating the LLM to generate forbidden content), there is significantly less discussion on application-level security and how to mitigate it. Thus, in this work we adopt an application-centric approach to GenAI security, and show that while LLMs cannot protect against ad-hoc application specific threats, they can provide the framework for applications to protect themselves against such threats. Our first contribution is defining Security Steerability - a novel security measure for LLMs, assessing the model's capability to adhere to strict guardrails that are defined in the system prompt ('Refrain from discussing about politics'). These guardrails, in case effective, can stop threats in the presence of malicious users who attempt to circumvent the application and cause harm to its providers. Our second contribution is a methodology to measure the security steerability of LLMs, utilizing two newly-developed datasets: VeganRibs assesses the LLM behavior in forcing specific guardrails that are not security per se in the presence of malicious user that uses attack boosters (jailbreaks and perturbations), and ReverseText takes this approach further and measures the LLM ability to force specific treatment of the user input as plain text while do user try to give it additional meanings...

We examine two properties of AI systems: capability (what a system can do) and steerability (how reliably one can shift behavior toward intended outcomes). A central question is whether capability growth reduces steerability and risks control collapse. We also distinguish between authorized steerability (builders reliably reaching intended behaviors) and unauthorized steerability (attackers eliciting disallowed behaviors). This distinction highlights a fundamental safety--security dilemma of AI models: safety requires high steerability to enforce control (e.g., stop/refuse), while security requires low steerability for malicious actors to elicit harmful behaviors. This tension presents a significant challenge for open-weight models, which currently exhibit high steerability via common techniques like fine-tuning or adversarial attacks. Using Qwen3 and InstrumentalEval, we find that a short anti-instrumental prompt suffix sharply reduces the measured convergence rate (e.g., shutdown avoidance, self-replication). For Qwen3-30B Instruct, the convergence rate drops from 81.69% under a pro-instrumental suffix to 2.82% under an anti-instrumental suffix. Under anti-instrumental prompting, larger aligned models show lower convergence rates than smaller ones (Instruct: 2.82% vs. 4.23%; Thinking: 4.23% vs. 9.86%). Code is available at github.com/j-hoscilowicz/instrumental_steering.

Despite significant progress in alignment, large language models (LLMs) remain vulnerable to adversarial attacks that elicit harmful behaviors. Activation steering techniques offer a promising inference-time intervention approach, but existing methods suffer from critical limitations: activation addition requires careful coefficient tuning and is sensitive to layer-specific norm variations, while directional ablation provides only binary control. Recent work on Angular Steering introduces continuous control via rotation in a 2D subspace, but its practical implementation violates norm preservation, causing distribution shift and generation collapse, particularly in models below 7B parameters. We propose Selective Steering, which addresses these limitations through two key innovations: (1) a mathematically rigorous norm-preserving rotation formulation that maintains activation distribution integrity, and (2) discriminative layer selection that applies steering only where feature representations exhibit opposite-signed class alignment. Experiments across nine models demonstrate that Selective Steering achieves 5.5x higher attack success rates than prior methods while maintaining zero perplexity violations and approximately 100\% capability retention on standard benchmarks. Our approach provides a principled, efficient framework for controllable and stable LLM behavior modification. Code: https://github.com/knoveleng/steering

Activation steering is a promising technique for controlling LLM behavior by adding semantically meaningful vectors directly into a model's hidden states during inference. It is often framed as a precise, interpretable, and potentially safer alternative to fine-tuning. We demonstrate the opposite: steering systematically breaks model alignment safeguards, making it comply with harmful requests. Through extensive experiments on different model families, we show that even steering in a random direction can increase the probability of harmful compliance from 0% to 2-27%. Alarmingly, steering benign features from a sparse autoencoder (SAE), a common source of interpretable directions, increases these rates by a further 2-4%. Finally, we show that combining 20 randomly sampled vectors that jailbreak a single prompt creates a universal attack, significantly increasing harmful compliance on unseen requests. These results challenge the paradigm of safety through interpretability, showing that precise control over model internals does not guarantee precise control over model behavior.

Mixture-of-Experts (MoE) in Large Language Models (LLMs) routes each token through a subset of specialized Feed-Forward Networks (FFN), known as experts. We present SteerMoE, a framework for steering MoE models by detecting and controlling behavior-linked experts. Our detection method identifies experts with distinct activation patterns across paired inputs exhibiting contrasting behaviors. By selectively (de)activating such experts during inference, we control behaviors like faithfulness and safety without retraining or modifying weights. Across 11 benchmarks and 6 LLMs, our steering raises safety by up to +20% and faithfulness by +27%. In adversarial attack mode, it drops safety by -41% alone, and -100% when combined with existing jailbreak methods, bypassing all safety guardrails and exposing a new dimension of alignment faking hidden within experts.

Current large language models have dangerous capabilities, which are likely to become more problematic in the future. Activation steering techniques can be used to reduce risks from these capabilities. In this paper, we investigate the efficacy of activation steering for broad skills and multiple behaviours. First, by comparing the effects of reducing performance on general coding ability and Python-specific ability, we find that steering broader skills is competitive to steering narrower skills. Second, we steer models to become more or less myopic and wealth-seeking, among other behaviours. In our experiments, combining steering vectors for multiple different behaviours into one steering vector is largely unsuccessful. On the other hand, injecting individual steering vectors at different places in a model simultaneously is promising.

We introduce SteeringControl, a benchmark for evaluating representation steering methods across core alignment objectives--bias, harmful generation, and hallucination--and their effects on secondary behaviors such as sycophancy and commonsense morality. While prior alignment work often highlights truthfulness or reasoning ability to demonstrate the side effects of representation steering, we find there are many unexplored tradeoffs not yet understood in a systematic way. We collect a dataset of safety-relevant primary and secondary behaviors to evaluate steering effectiveness and behavioral entanglement centered around five popular steering methods. To enable this, we craft a modular steering framework based on unique components that serve as the building blocks of many existing methods. Our results on Qwen-2.5-7B and Llama-3.1-8B find that strong steering performance is dependent on the specific combination of steering method, model, and targeted behavior, and that severe concept entanglement can result from poor combinations of these three as well. We release our code here: https://github.com/wang-research-lab/SteeringControl.git.

Natural-language user profiles have recently attracted attention not only for improved interpretability, but also for their potential to make recommender systems more steerable. By enabling direct editing, natural-language profiles allow users to explicitly articulate preferences that may be difficult to infer from past behavior. However, it remains unclear whether current natural-language-based recommendation methods can follow such steering commands. While existing steerability evaluations have shown some success for well-recognized item attributes (e.g., movie genres), we argue that these benchmarks fail to capture the richer forms of user control that motivate steerable recommendations. To address this gap, we introduce SteerEval, an evaluation framework designed to measure more nuanced and diverse forms of steerability by using interventions that range from genres to content-warning for movies. We assess the steerability of a family of pretrained natural-language recommenders, examine the potential and limitations of steering on relatively niche topics, and compare how different profile and recommendation interventions impact steering effectiveness. Finally, we offer practical design suggestions informed by our findings and discuss future steps in steerable recommender design.

Despite advances in large language models (LLMs) on reasoning and instruction-following benchmarks, it remains unclear whether they can reliably produce outputs aligned with a broad variety of user goals, a concept we refer to as steerability. The abundance of methods proposed to modify LLM behavior makes it unclear whether current LLMs are already steerable, or require further intervention. In particular, LLMs may exhibit (i) poor coverage, where rare user goals are underrepresented; (ii) miscalibration, where models overshoot requests; and (iii) side effects, where changes to one dimension of text inadvertently affect others. To systematically evaluate these failures, we introduce a framework based on a multi-dimensional goal space that models user goals and LLM outputs as vectors with dimensions corresponding to text attributes (e.g., reading difficulty). Applied to a text-rewriting task, we find that current LLMs struggle with steerability, as side effects are persistent. Interventions to improve steerability, such as prompt engineering, best-of-N sampling, and reinforcement learning fine-tuning, have varying effectiveness, yet side effects remain problematic. Our findings suggest that even strong LLMs struggle with steerability, and existing alignment strategies may be insufficient. We open-source our steerability evaluation framework at https://github.com/MLD3/steerability.

Steering methods influence Large Language Model behavior by identifying semantic directions in hidden representations, but are typically realized through inference-time activation interventions that apply a fixed, global modification to the model's internal states. While effective, such interventions often induce unfavorable attribute-utility trade-offs under strong control, as they ignore the fact that many behaviors are governed by a small and heterogeneous subset of model components. We propose Steer2Edit, a theoretically grounded, training-free framework that transforms steering vectors from inference-time control signals into diagnostic signals for component-level rank-1 weight editing. Instead of uniformly injecting a steering direction during generation, Steer2Edit selectively redistributes behavioral influence across individual attention heads and MLP neurons, yielding interpretable edits that preserve the standard forward pass and remain compatible with optimized parallel inference. Across safety alignment, hallucination mitigation, and reasoning efficiency, Steer2Edit consistently achieves more favorable attribute-utility trade-offs: at matched downstream performance, it improves safety by up to 17.2%, increases truthfulness by 9.8%, and reduces reasoning length by 12.2% on average. Overall, Steer2Edit provides a principled bridge between representation steering and weight editing by translating steering signals into interpretable, training-free parameter updates.

Activation steering has emerged as a promising approach for efficiently adapting large language models (LLMs) to downstream behaviors. However, most existing steering methods rely on a single static direction per task or concept, making them inflexible under task variation and inadequate for complex tasks that require multiple coordinated capabilities. To address this limitation, we propose STEER2ADAPT, a lightweight framework that adapts LLMs by composing steering vectors rather than learning new ones from scratch. In many domains (e.g., reasoning or safety), tasks share a small set of underlying concept dimensions. STEER2ADAPT captures these dimensions as a reusable, low-dimensional semantic prior subspace, and adapts to new tasks by dynamically discovering a linear combination of basis vectors from only a handful of examples. Experiments across 9 tasks and 3 models in both reasoning and safety domains demonstrate the effectiveness of STEER2ADAPT, achieving an average improvement of 8.2%. Extensive analyses further show that STEER2ADAPT is a data-efficient, stable, and transparent inference-time adaptation method for LLMs.

Large language models (LLMs) show inherent brittleness in their safety mechanisms, as evidenced by their susceptibility to jailbreaking and even non-malicious fine-tuning. This study explores this brittleness of safety alignment by leveraging pruning and low-rank modifications. We develop methods to identify critical regions that are vital for safety guardrails, and that are disentangled from utility-relevant regions at both the neuron and rank levels. Surprisingly, the isolated regions we find are sparse, comprising about 3% at the parameter level and 2.5% at the rank level. Removing these regions compromises safety without significantly impacting utility, corroborating the inherent brittleness of the model's safety mechanisms. Moreover, we show that LLMs remain vulnerable to low-cost fine-tuning attacks even when modifications to the safety-critical regions are restricted. These findings underscore the urgent need for more robust safety strategies in LLMs.

Precise control over language model generation is vital for ensuring both safety and reliability. Although prompt engineering and steering are commonly used to intervene in model behaviors, the vast number of parameters in models often results in highly intertwined internal representations. This interdependency can limit control precision and sometimes lead to unintended side effects. Recent research has explored the use of sparse autoencoders (SAE) to disentangle knowledge in high-dimensional spaces for steering. However, these applications have been limited to toy tasks owing to the nontrivial issue of locating atomic knowledge components. In this paper, we propose Steering Target Atoms (STA), a novel method that isolates and manipulates disentangled knowledge components to enhance safety. Comprehensive experiments demonstrate the effectiveness of our approach. Further analysis reveals that steering exhibits superior robustness and flexibility, particularly in adversarial scenarios. We also apply the steering strategy to the large reasoning model, confirming its effectiveness in precise reasoning control.

To ensure AI safety, instruction-tuned Large Language Models (LLMs) are specifically trained to ensure alignment, which refers to making models behave in accordance with human intentions. While these models have demonstrated commendable results on various safety benchmarks, the vulnerability of their safety alignment has not been extensively studied. This is particularly troubling given the potential harm that LLMs can inflict. Existing attack methods on LLMs often rely on poisoned training data or the injection of malicious prompts. These approaches compromise the stealthiness and generalizability of the attacks, making them susceptible to detection. Additionally, these models often demand substantial computational resources for implementation, making them less practical for real-world applications. Inspired by recent success in modifying model behavior through steering vectors without the need for optimization, and drawing on its effectiveness in red-teaming LLMs, we conducted experiments employing activation steering to target four key aspects of LLMs: truthfulness, toxicity, bias, and harmfulness - across a varied set of attack settings. To establish a universal attack strategy applicable to diverse target alignments without depending on manual analysis, we automatically select the intervention layer based on contrastive layer search. Our experiment results show that activation attacks are highly effective and add little or no overhead to attack efficiency. Additionally, we discuss potential countermeasures against such activation attacks. Our code and data are available at https://github.com/wang2226/Backdoor-Activation-Attack Warning: this paper contains content that can be offensive or upsetting.

As large language models are deployed in high-stakes enterprise applications, from healthcare to finance, ensuring adherence to organization-specific policies has become essential. Yet existing safety evaluations focus exclusively on universal harms. We present COMPASS (Company/Organization Policy Alignment Assessment), the first systematic framework for evaluating whether LLMs comply with organizational allowlist and denylist policies. We apply COMPASS to eight diverse industry scenarios, generating and validating 5,920 queries that test both routine compliance and adversarial robustness through strategically designed edge cases. Evaluating seven state-of-the-art models, we uncover a fundamental asymmetry: models reliably handle legitimate requests (>95% accuracy) but catastrophically fail at enforcing prohibitions, refusing only 13-40% of adversarial denylist violations. These results demonstrate that current LLMs lack the robustness required for policy-critical deployments, establishing COMPASS as an essential evaluation framework for organizational AI safety.

Recent work by Anthropic on Mechanistic interpretability claims to understand and control Large Language Models by extracting human-interpretable features from their neural activation patterns using sparse autoencoders (SAEs). If successful, this approach offers one of the most promising routes for human oversight in AI safety. We conduct an initial stress-test of these claims by replicating their main results with open-source SAEs for Llama 3.1. While we successfully reproduce basic feature extraction and steering capabilities, our investigation suggests that major caution is warranted regarding the generalizability of these claims. We find that feature steering exhibits substantial fragility, with sensitivity to layer selection, steering magnitude, and context. We observe non-standard activation behavior and demonstrate the difficulty to distinguish thematically similar features from one another. While SAE-based interpretability produces compelling demonstrations in selected cases, current methods often fall short of the systematic reliability required for safety-critical applications. This suggests a necessary shift in focus from prioritizing interpretability of internal representations toward reliable prediction and control of model output. Our work contributes to a more nuanced understanding of what mechanistic interpretability has achieved and highlights fundamental challenges for AI safety that remain unresolved.

Responsible practices for deploying language models include guiding models to recognize and refuse answering prompts that are considered unsafe, while complying with safe prompts. Achieving such behavior typically requires updating model weights, which is costly and inflexible. We explore opportunities to steering model activations at inference time, which does not require updating weights. Using sparse autoencoders, we identify and steer features in Phi-3 Mini that mediate refusal behavior. We find that feature steering can improve Phi-3 Minis robustness to jailbreak attempts across various harms, including challenging multi-turn attacks. However, we discover that feature steering can adversely affect overall performance on benchmarks. These results suggest that identifying steerable mechanisms for refusal via sparse autoencoders is a promising approach for enhancing language model safety, but that more research is needed to mitigate feature steerings adverse effects on performance.

We introduce Refusal Steering, an inference-time method to exercise fine-grained control over Large Language Models refusal behaviour on politically sensitive topics without retraining. We replace fragile pattern-based refusal detection with an LLM-as-a-judge that assigns refusal confidence scores and we propose a ridge-regularized variant to compute steering vectors that better isolate the refusal--compliance direction. On Qwen3-Next-80B-A3B-Thinking, our method removes the refusal behaviour of the model around politically sensitive topics while maintaining safety on JailbreakBench and near-baseline performance on general benchmarks. The approach generalizes across 4B and 80B models and can also induce targeted refusals when desired. We analize the steering vectors and show that refusal signals concentrate in deeper layers of the transformer and are distributed across many dimensions. Together, these results demonstrate that activation steering can remove political refusal behaviour while retaining safety alignment for harmful content, offering a practical path to controllable, transparent moderation at inference time.

Recent advancements in large vision language models (VLMs) tailored for autonomous driving (AD) have shown strong scene understanding and reasoning capabilities, making them undeniable candidates for end-to-end driving systems. However, limited work exists on studying the trustworthiness of DriveVLMs -- a critical factor that directly impacts public transportation safety. In this paper, we introduce AutoTrust, a comprehensive trustworthiness benchmark for large vision-language models in autonomous driving (DriveVLMs), considering diverse perspectives -- including trustfulness, safety, robustness, privacy, and fairness. We constructed the largest visual question-answering dataset for investigating trustworthiness issues in driving scenarios, comprising over 10k unique scenes and 18k queries. We evaluated six publicly available VLMs, spanning from generalist to specialist, from open-source to commercial models. Our exhaustive evaluations have unveiled previously undiscovered vulnerabilities of DriveVLMs to trustworthiness threats. Specifically, we found that the general VLMs like LLaVA-v1.6 and GPT-4o-mini surprisingly outperform specialized models fine-tuned for driving in terms of overall trustworthiness. DriveVLMs like DriveLM-Agent are particularly vulnerable to disclosing sensitive information. Additionally, both generalist and specialist VLMs remain susceptible to adversarial attacks and struggle to ensure unbiased decision-making across diverse environments and populations. Our findings call for immediate and decisive action to address the trustworthiness of DriveVLMs -- an issue of critical importance to public safety and the welfare of all citizens relying on autonomous transportation systems. Our benchmark is publicly available at https://github.com/taco-group/AutoTrust, and the leaderboard is released at https://taco-group.github.io/AutoTrust/.

Intervention-based model steering offers a lightweight and interpretable alternative to prompting and fine-tuning. However, by adapting strong optimization objectives from fine-tuning, current methods are susceptible to overfitting and often underperform, sometimes generating unnatural outputs. We hypothesize that this is because effective steering requires the faithful identification of internal model mechanisms, not the enforcement of external preferences. To this end, we build on the principles of distributed alignment search (DAS), the standard for causal variable localization, to propose a new steering method: Concept DAS (CDAS). While we adopt the core mechanism of DAS, distributed interchange intervention (DII), we introduce a novel distribution matching objective tailored for the steering task by aligning intervened output distributions with counterfactual distributions. CDAS differs from prior work in two main ways: first, it learns interventions via weak-supervised distribution matching rather than probability maximization; second, it uses DIIs that naturally enable bi-directional steering and allow steering factors to be derived from data, reducing the effort required for hyperparameter tuning and resulting in more faithful and stable control. On AxBench, a large-scale model steering benchmark, we show that CDAS does not always outperform preference-optimization methods but may benefit more from increased model scale. In two safety-related case studies, overriding refusal behaviors of safety-aligned models and neutralizing a chain-of-thought backdoor, CDAS achieves systematic steering while maintaining general model utility. These results indicate that CDAS is complementary to preference-optimization approaches and conditionally constitutes a robust approach to intervention-based model steering. Our code is available at https://github.com/colored-dye/concept_das.

Large language models can resist task-misaligned activation steering during inference, sometimes recovering mid-generation to produce improved responses even when steering remains active. We term this Endogenous Steering Resistance (ESR). Using sparse autoencoder (SAE) latents to steer model activations, we find that Llama-3.3-70B shows substantial ESR, while smaller models from the Llama-3 and Gemma-2 families exhibit the phenomenon less frequently. We identify 26 SAE latents that activate differentially during off-topic content and are causally linked to ESR in Llama-3.3-70B. Zero-ablating these latents reduces the multi-attempt rate by 25%, providing causal evidence for dedicated internal consistency-checking circuits. We demonstrate that ESR can be deliberately enhanced through both prompting and training: meta-prompts instructing the model to self-monitor increase the multi-attempt rate by 4x for Llama-3.3-70B, and fine-tuning on self-correction examples successfully induces ESR-like behavior in smaller models. These findings have dual implications: ESR could protect against adversarial manipulation but might also interfere with beneficial safety interventions that rely on activation steering. Understanding and controlling these resistance mechanisms is important for developing transparent and controllable AI systems. Code is available at github.com/agencyenterprise/endogenous-steering-resistance.

Analysis of learned representations has a blind spot: it focuses on similarity, measuring how closely embeddings align with external references, but similarity reveals only what is represented, not whether that structure is robust. We introduce geometric stability, a distinct dimension that quantifies how reliably representational geometry holds under perturbation, and present Shesha, a framework for measuring it. Across 2,463 configurations in seven domains, we show that stability and similarity are empirically uncorrelated (ρapprox 0.01) and mechanistically distinct: similarity metrics collapse after removing the top principal components, while stability retains sensitivity to fine-grained manifold structure. This distinction yields actionable insights: for safety monitoring, stability acts as a functional geometric canary, detecting structural drift nearly 2times more sensitively than CKA while filtering out the non-functional noise that triggers false alarms in rigid distance metrics; for controllability, supervised stability predicts linear steerability (ρ= 0.89-0.96); for model selection, stability dissociates from transferability, revealing a geometric tax that transfer optimization incurs. Beyond machine learning, stability predicts CRISPR perturbation coherence and neural-behavioral coupling. By quantifying how reliably systems maintain structure, geometric stability provides a necessary complement to similarity for auditing representations across biological and computational systems.

Recent advances in Large Language Models (LLMs) have sparked concerns over their potential to acquire and misuse dangerous or high-risk capabilities, posing frontier risks. Current safety evaluations primarily test for what a model can do - its capabilities - without assessing what it would do if endowed with high-risk capabilities. This leaves a critical blind spot: models may strategically conceal capabilities or rapidly acquire them, while harboring latent inclinations toward misuse. We argue that propensity - the likelihood of a model to pursue harmful actions if empowered - is a critical, yet underexplored, axis of safety evaluation. We present PropensityBench, a novel benchmark framework that assesses the proclivity of models to engage in risky behaviors when equipped with simulated dangerous capabilities using proxy tools. Our framework includes 5,874 scenarios with 6,648 tools spanning four high-risk domains: cybersecurity, self-proliferation, biosecurity, and chemical security. We simulate access to powerful capabilities via a controlled agentic environment and evaluate the models' choices under varying operational pressures that reflect real-world constraints or incentives models may encounter, such as resource scarcity or gaining more autonomy. Across open-source and proprietary frontier models, we uncover 9 alarming signs of propensity: models frequently choose high-risk tools when under pressure, despite lacking the capability to execute such actions unaided. These findings call for a shift from static capability audits toward dynamic propensity assessments as a prerequisite for deploying frontier AI systems safely. Our code is available at https://github.com/scaleapi/propensity-evaluation.

Large language models (LLMs) can sometimes detect when they are being evaluated and adjust their behavior to appear more aligned, compromising the reliability of safety evaluations. In this paper, we show that adding a steering vector to an LLM's activations can suppress evaluation-awareness and make the model act like it is deployed during evaluation. To study our steering technique, we train an LLM to exhibit evaluation-aware behavior using a two-step training process designed to mimic how this behavior could emerge naturally. First, we perform continued pretraining on documents with factual descriptions of the model (1) using Python type hints during evaluation but not during deployment and (2) recognizing that the presence of a certain evaluation cue always means that it is being tested. Then, we train the model with expert iteration to use Python type hints in evaluation settings. The resulting model is evaluation-aware: it writes type hints in evaluation contexts more than deployment contexts. We find that activation steering can suppress evaluation awareness and make the model act like it is deployed even when the cue is present. Importantly, we constructed our steering vector using the original model before our additional training. Our results suggest that AI evaluators could improve the reliability of safety evaluations by steering models to act like they are deployed.

Modern deployments require LLMs to enforce safety policies at scale, yet many controls rely on inference-time interventions that add recurring compute cost and serving complexity. Activation steering is widely used, but it requires runtime hooks and scales cost with the number of generations; conditional variants improve selectivity by gating when steering is applied but still retain an inference-time control path. We ask whether selective refusal can be moved entirely offline: can a mechanistic understanding of category-specific refusal be distilled into a circuit-restricted weight update that deploys as a standard checkpoint? We propose C-Δθ: Circuit Restricted Weight Arithmetic, which (i) localizes refusal-causal computation as a sparse circuit using EAP-IG and (ii) computes a constrained weight update ΔθC supported only on that circuit (typically <5% of parameters). Applying ΔθC yields a drop-in edited checkpoint with no inference-time hooks, shifting cost from per-request intervention to a one-time offline update. We evaluate category-targeted selectivity and capability retention on refusal and utility benchmarks.

While Reinforcement Learning with Verifiable Rewards (RLVR) can enhance LLM reasoning, its training process poses a critical risk: entropy collapse. This phenomenon is a rapid loss of policy diversity, stemming from the exploration-exploitation imbalance and leading to a lack of generalization. Recent entropy-intervention methods aim to prevent entropy collapse, yet their underlying mechanisms remain unclear. In this paper, we conduct a quantitative analysis to reveal token-level entropy changes and how existing entropy intervention methods help avoid entropy collapse. Our findings point out a fundamental limitation of existing methods: they attempt to control entropy dynamics indirectly. By only affecting related factors, such as the advantage signal and generation probability, their effectiveness is inherently limited and could potentially fail. To address this limitation, we introduce an entropy-change-aware reweighting scheme, namely Stabilizing Token-level Entropy-changE via Reweighting (STEER), that adaptively stabilizes entropy dynamics through fine-grained token-level adjustments. Our approach mitigates over-exploitation while fostering robust exploration. Extensive experiments demonstrate that STEER significantly mitigates entropy collapse, stabilizes entropy dynamics, and achieves stronger downstream performance across various mathematical reasoning benchmarks \footnote{Our code is available at https://github.com/zz-haooo/STEER.

Large language models (LLMs) typically deploy safety mechanisms to prevent harmful content generation. Most current approaches focus narrowly on risks posed by malicious actors, often framing risks as adversarial events and relying on defensive refusals. However, in real-world settings, risks also come from non-malicious users seeking help while under psychological distress (e.g., self-harm intentions). In such cases, the model's response can strongly influence the user's next actions. Simple refusals may lead them to repeat, escalate, or move to unsafe platforms, creating worse outcomes. We introduce Constructive Safety Alignment (CSA), a human-centric paradigm that protects against malicious misuse while actively guiding vulnerable users toward safe and helpful results. Implemented in Oyster-I (Oy1), CSA combines game-theoretic anticipation of user reactions, fine-grained risk boundary discovery, and interpretable reasoning control, turning safety into a trust-building process. Oy1 achieves state-of-the-art safety among open models while retaining high general capabilities. On our Constructive Benchmark, it shows strong constructive engagement, close to GPT-5, and unmatched robustness on the Strata-Sword jailbreak dataset, nearing GPT-o1 levels. By shifting from refusal-first to guidance-first safety, CSA redefines the model-user relationship, aiming for systems that are not just safe, but meaningfully helpful. We release Oy1, code, and the benchmark to support responsible, user-centered AI.

Emerging multi-robot systems rely on cooperation between humans and robots, with robots following automatically generated motion plans to service application-level tasks. Given the safety requirements associated with operating in proximity to humans and expensive infrastructure, it is important to understand and mitigate the security vulnerabilities of such systems caused by compromised robots who diverge from their assigned plans. We focus on centralized systems, where a *central entity* (CE) is responsible for determining and transmitting the motion plans to the robots, which report their location as they move following the plan. The CE checks that robots follow their assigned plans by comparing their expected location to the location they self-report. We show that this self-reporting monitoring mechanism is vulnerable to *plan-deviation attacks* where compromised robots don't follow their assigned plans while trying to conceal their movement by mis-reporting their location. We propose a two-pronged mitigation for plan-deviation attacks: (1) an attack detection technique leveraging both the robots' local sensing capabilities to report observations of other robots and *co-observation schedules* generated by the CE, and (2) a prevention technique where the CE issues *horizon-limiting announcements* to the robots, reducing their instantaneous knowledge of forward lookahead steps in the global motion plan. On a large-scale automated warehouse benchmark, we show that our solution enables attack prevention guarantees from a stealthy attacker that has compromised multiple robots.

Recent developments in Large Language Models (LLMs) have manifested significant advancements. To facilitate safeguards against malicious exploitation, a body of research has concentrated on aligning LLMs with human preferences and inhibiting their generation of inappropriate content. Unfortunately, such alignments are often vulnerable: fine-tuning with a minimal amount of harmful data can easily unalign the target LLM. While being effective, such fine-tuning-based unalignment approaches also have their own limitations: (1) non-stealthiness, after fine-tuning, safety audits or red-teaming can easily expose the potential weaknesses of the unaligned models, thereby precluding their release/use. (2) non-persistence, the unaligned LLMs can be easily repaired through re-alignment, i.e., fine-tuning again with aligned data points. In this work, we show that it is possible to conduct stealthy and persistent unalignment on large language models via backdoor injections. We also provide a novel understanding on the relationship between the backdoor persistence and the activation pattern and further provide guidelines for potential trigger design. Through extensive experiments, we demonstrate that our proposed stealthy and persistent unalignment can successfully pass the safety evaluation while maintaining strong persistence against re-alignment defense.

Deploying large, complex policies in the real world requires the ability to steer them to fit the needs of a situation. Most common steering approaches, like goal-conditioning, require training the robot policy with a distribution of test-time objectives in mind. To overcome this limitation, we present DynaGuide, a steering method for diffusion policies using guidance from an external dynamics model during the diffusion denoising process. DynaGuide separates the dynamics model from the base policy, which gives it multiple advantages, including the ability to steer towards multiple objectives, enhance underrepresented base policy behaviors, and maintain robustness on low-quality objectives. The separate guidance signal also allows DynaGuide to work with off-the-shelf pretrained diffusion policies. We demonstrate the performance and features of DynaGuide against other steering approaches in a series of simulated and real experiments, showing an average steering success of 70% on a set of articulated CALVIN tasks and outperforming goal-conditioning by 5.4x when steered with low-quality objectives. We also successfully steer an off-the-shelf real robot policy to express preference for particular objects and even create novel behavior. Videos and more can be found on the project website: https://dynaguide.github.io

Despite investments in improving model safety, studies show that misaligned capabilities remain latent in safety-tuned models. In this work, we shed light on the mechanics of this phenomenon. First, we show that even when model generations are safe, harmful content can persist in hidden representations and can be extracted by decoding from earlier layers. Then, we show that whether the model divulges such content depends significantly on its perception of who it is talking to, which we refer to as user persona. In fact, we find manipulating user persona to be even more effective for eliciting harmful content than direct attempts to control model refusal. We study both natural language prompting and activation steering as control methods and show that activation steering is significantly more effective at bypassing safety filters. We investigate why certain personas break model safeguards and find that they enable the model to form more charitable interpretations of otherwise dangerous queries. Finally, we show we can predict a persona's effect on refusal given only the geometry of its steering vector.

Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks - ranging from harmful content generation to broader societal harms - pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering - simple vector arithmetic for steering model's behavior during inference - to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities.

Visual-Language-Action (VLA) models report impressive success rates on robotic manipulation benchmarks, yet these results may mask fundamental weaknesses in robustness. We perform a systematic vulnerability analysis by introducing controlled perturbations across seven dimensions: objects layout, camera viewpoints, robot initial states, language instructions, light conditions, background textures and sensor noise. We comprehensively analyzed multiple state-of-the-art models and revealed consistent brittleness beneath apparent competence. Our analysis exposes critical weaknesses: models exhibit extreme sensitivity to perturbation factors, including camera viewpoints and robot initial states, with performance dropping from 95% to below 30% under modest perturbations. Surprisingly, models are largely insensitive to language variations, with further experiments revealing that models tend to ignore language instructions completely. Our findings challenge the assumption that high benchmark scores equate to true competency and highlight the need for evaluation practices that assess reliability under realistic variation.

Vision-Language-Action (VLA) models have advanced robotic control by enabling end-to-end decision-making directly from multimodal inputs. However, their tightly coupled architectures expose novel security vulnerabilities. Unlike traditional adversarial perturbations, backdoor attacks represent a stealthier, persistent, and practically significant threat-particularly under the emerging Training-as-a-Service paradigm-but remain largely unexplored in the context of VLA models. To address this gap, we propose BadVLA, a backdoor attack method based on Objective-Decoupled Optimization, which for the first time exposes the backdoor vulnerabilities of VLA models. Specifically, it consists of a two-stage process: (1) explicit feature-space separation to isolate trigger representations from benign inputs, and (2) conditional control deviations that activate only in the presence of the trigger, while preserving clean-task performance. Empirical results on multiple VLA benchmarks demonstrate that BadVLA consistently achieves near-100% attack success rates with minimal impact on clean task accuracy. Further analyses confirm its robustness against common input perturbations, task transfers, and model fine-tuning, underscoring critical security vulnerabilities in current VLA deployments. Our work offers the first systematic investigation of backdoor vulnerabilities in VLA models, highlighting an urgent need for secure and trustworthy embodied model design practices. We have released the project page at https://badvla-project.github.io/.

As vision-language models (VLMs) gain prominence, their multimodal interfaces also introduce new safety vulnerabilities, making the safety evaluation challenging and critical. Existing red-teaming efforts are either restricted to a narrow set of adversarial patterns or depend heavily on manual engineering, lacking scalable exploration of emerging real-world VLM vulnerabilities. To bridge this gap, we propose ARMs, an adaptive red-teaming agent that systematically conducts comprehensive risk assessments for VLMs. Given a target harmful behavior or risk definition, ARMs automatically optimizes diverse red-teaming strategies with reasoning-enhanced multi-step orchestration, to effectively elicit harmful outputs from target VLMs. We propose 11 novel multimodal attack strategies, covering diverse adversarial patterns of VLMs (e.g., reasoning hijacking, contextual cloaking), and integrate 17 red-teaming algorithms into ARMs via model context protocol (MCP). To balance the diversity and effectiveness of the attack, we design a layered memory with an epsilon-greedy attack exploration algorithm. Extensive experiments on instance- and policy-based benchmarks show that ARMs achieves SOTA attack success rates, exceeding baselines by an average of 52.1% and surpassing 90% on Claude-4-Sonnet. We show that the diversity of red-teaming instances generated by ARMs is significantly higher, revealing emerging vulnerabilities in VLMs. Leveraging ARMs, we construct ARMs-Bench, a large-scale multimodal safety dataset comprising over 30K red-teaming instances spanning 51 diverse risk categories, grounded in both real-world multimodal threats and regulatory risks. Safety fine-tuning with ARMs-Bench substantially improves the robustness of VLMs while preserving their general utility, providing actionable guidance to improve multimodal safety alignment against emerging threats.

Despite the state-of-the-art performance of large language models (LLMs) across diverse tasks, their susceptibility to adversarial attacks and unsafe content generation remains a major obstacle to deployment, particularly in high-stakes settings. Addressing this challenge requires safety mechanisms that are both practically effective and supported by rigorous theory. We introduce BarrierSteer, a novel framework that formalizes response safety by embedding learned non-linear safety constraints directly into the model's latent representation space. BarrierSteer employs a steering mechanism based on Control Barrier Functions (CBFs) to efficiently detect and prevent unsafe response trajectories during inference with high precision. By enforcing multiple safety constraints through efficient constraint merging, without modifying the underlying LLM parameters, BarrierSteer preserves the model's original capabilities and performance. We provide theoretical results establishing that applying CBFs in latent space offers a principled and computationally efficient approach to enforcing safety. Our experiments across multiple models and datasets show that BarrierSteer substantially reduces adversarial success rates, decreases unsafe generations, and outperforms existing methods.

Safety alignment in Large Language Models (LLMs) often involves mediating internal representations to refuse harmful requests. Recent research has demonstrated that these safety mechanisms can be bypassed by ablating or removing specific representational directions within the model. In this paper, we propose the opposite approach: Rank-One Safety Injection (ROSI), a white-box method that amplifies a model's safety alignment by permanently steering its activations toward the refusal-mediating subspace. ROSI operates as a simple, fine-tuning-free rank-one weight modification applied to all residual stream write matrices. The required safety direction can be computed from a small set of harmful and harmless instruction pairs. We show that ROSI consistently increases safety refusal rates - as evaluated by Llama Guard 3 - while preserving the utility of the model on standard benchmarks such as MMLU, HellaSwag, and Arc. Furthermore, we show that ROSI can also re-align 'uncensored' models by amplifying their own latent safety directions, demonstrating its utility as an effective last-mile safety procedure. Our results suggest that targeted, interpretable weight steering is a cheap and potent mechanism to improve LLM safety, complementing more resource-intensive fine-tuning paradigms.

Controlling undesirable Large Language Model (LLM) behaviors, such as the generation of unsafe content or failing to adhere to safety guidelines, often relies on costly fine-tuning. Activation steering provides an alternative for inference-time control, but existing methods typically lack fine-grained, adaptive mechanisms. We introduce a novel approach using a lightweight, trainable controller network integrated during inference. This controller network observes specific intermediate LLM activations and predicts both a global scaling factor and layer-specific weights. The predicted global scaling factor and layer-specific weights then dynamically modulate the intensity of a steering patch, derived from a pre-computed "refusal direction" vector, applied across the LLM's layers during generation. Trained on activations from both harmful and benign prompts, our controller learns to discriminatively apply nuanced, layer-aware interventions, activating steering primarily for harmful inputs. Experiments using safety benchmarks like ToxicChat & In-The-Wild Jailbreak Prompts demonstrate that our weighted steering controller significantly increases refusal rates compared to the base LLM, achieving targeted behavioral modification without altering the original model parameters. Our experiments with Llama-3.1-8B, Llama-3.2-1B & Mistral-7B show our approach outperforms existing methods, presenting an efficient and adaptive method for fine-grained control over LLM behavior at inference time.

Recent progress in image-to-3D has opened up immense possibilities for design, AR/VR, and robotics. However, to use AI-generated 3D assets in real applications, a critical requirement is the capability to edit them easily. We present a feedforward method, Steer3D, to add text steerability to image-to-3D models, which enables editing of generated 3D assets with language. Our approach is inspired by ControlNet, which we adapt to image-to-3D generation to enable text steering directly in a forward pass. We build a scalable data engine for automatic data generation, and develop a two-stage training recipe based on flow-matching training and Direct Preference Optimization (DPO). Compared to competing methods, Steer3D more faithfully follows the language instruction and maintains better consistency with the original 3D asset, while being 2.4x to 28.5x faster. Steer3D demonstrates that it is possible to add a new modality (text) to steer the generation of pretrained image-to-3D generative models with 100k data. Project website: https://glab-caltech.github.io/steer3d/

To control the behavior of language models, steering methods attempt to ensure that outputs of the model satisfy specific pre-defined properties. Adding steering vectors to the model is a promising method of model control that is easier than finetuning, and may be more robust than prompting. However, it can be difficult to anticipate the effects of steering vectors produced by almost all existing methods, such as CAA (Panickssery et al., 2024) or the direct use of SAE latents (Templeton et al., 2024). In our work, we address this issue by using SAEs to measure the effects of steering vectors, giving us a method that can be used to understand the causal effect of any steering vector intervention. We use this method for measuring causal effects to develop an improved steering method, SAE-Targeted Steering (SAE-TS), which finds steering vectors to target specific SAE features while minimizing unintended side effects. We show that overall, SAE-TS balances steering effects with coherence better than CAA and SAE feature steering, when evaluated on a range of tasks.

Evaluations of large language model (LLM) risks and capabilities are increasingly being incorporated into AI risk management and governance frameworks. Currently, most risk evaluations are conducted by designing inputs that elicit harmful behaviors from the system. However, a fundamental limitation of this approach is that the harmfulness of the behaviors identified during any particular evaluation can only lower bound the model's worst-possible-case behavior. As a complementary method for eliciting harmful behaviors, we propose evaluating LLMs with model tampering attacks which allow for modifications to latent activations or weights. We pit state-of-the-art techniques for removing harmful LLM capabilities against a suite of 5 input-space and 6 model tampering attacks. In addition to benchmarking these methods against each other, we show that (1) model resilience to capability elicitation attacks lies on a low-dimensional robustness subspace; (2) the attack success rate of model tampering attacks can empirically predict and offer conservative estimates for the success of held-out input-space attacks; and (3) state-of-the-art unlearning methods can easily be undone within 16 steps of fine-tuning. Together these results highlight the difficulty of removing harmful LLM capabilities and show that model tampering attacks enable substantially more rigorous evaluations than input-space attacks alone. We release models at https://huggingface.co/LLM-GAT

In autonomous driving (AD), accurate perception is indispensable to achieving safe and secure driving. Due to its safety-criticality, the security of AD perception has been widely studied. Among different attacks on AD perception, the physical adversarial object evasion attacks are especially severe. However, we find that all existing literature only evaluates their attack effect at the targeted AI component level but not at the system level, i.e., with the entire system semantics and context such as the full AD pipeline. Thereby, this raises a critical research question: can these existing researches effectively achieve system-level attack effects (e.g., traffic rule violations) in the real-world AD context? In this work, we conduct the first measurement study on whether and how effectively the existing designs can lead to system-level effects, especially for the STOP sign-evasion attacks due to their popularity and severity. Our evaluation results show that all the representative prior works cannot achieve any system-level effects. We observe two design limitations in the prior works: 1) physical model-inconsistent object size distribution in pixel sampling and 2) lack of vehicle plant model and AD system model consideration. Then, we propose SysAdv, a novel system-driven attack design in the AD context and our evaluation results show that the system-level effects can be significantly improved, i.e., the violation rate increases by around 70%.

As large language models (LLMs) become increasingly capable, it is prudent to assess whether safety measures remain effective even if LLMs intentionally try to bypass them. Previous work introduced control evaluations, an adversarial framework for testing deployment strategies of untrusted models (i.e., models which might be trying to bypass safety measures). While prior work treats a single failure as unacceptable, we perform control evaluations in a "distributed threat setting" -- a setting where no single action is catastrophic and no single action provides overwhelming evidence of misalignment. We approach this problem with a two-level deployment framework that uses an adaptive macro-protocol to choose between micro-protocols. Micro-protocols operate on a single task, using a less capable, but extensively tested (trusted) model to harness and monitor the untrusted model. Meanwhile, the macro-protocol maintains an adaptive credence on the untrusted model's alignment based on its past actions, using it to pick between safer and riskier micro-protocols. We evaluate our method in a code generation testbed where a red team attempts to generate subtly backdoored code with an LLM whose deployment is safeguarded by a blue team. We plot Pareto frontiers of safety (# of non-backdoored solutions) and usefulness (# of correct solutions). At a given level of usefulness, our adaptive deployment strategy reduces the number of backdoors by 80% compared to non-adaptive baselines.

The growing misuse of Vision-Language Models (VLMs) has led providers to deploy multiple safeguards, including alignment tuning, system prompts, and content moderation. However, the real-world robustness of these defenses against adversarial attacks remains underexplored. We introduce Multi-Faceted Attack (MFA), a framework that systematically exposes general safety vulnerabilities in leading defense-equipped VLMs such as GPT-4o, Gemini-Pro, and Llama-4. The core component of MFA is the Attention-Transfer Attack (ATA), which hides harmful instructions inside a meta task with competing objectives. We provide a theoretical perspective based on reward hacking to explain why this attack succeeds. To improve cross-model transferability, we further introduce a lightweight transfer-enhancement algorithm combined with a simple repetition strategy that jointly bypasses both input-level and output-level filters without model-specific fine-tuning. Empirically, we show that adversarial images optimized for one vision encoder transfer broadly to unseen VLMs, indicating that shared visual representations create a cross-model safety vulnerability. Overall, MFA achieves a 58.5% success rate and consistently outperforms existing methods. On state-of-the-art commercial models, MFA reaches a 52.8% success rate, surpassing the second-best attack by 34%. These results challenge the perceived robustness of current defense mechanisms and highlight persistent safety weaknesses in modern VLMs. Code: https://github.com/cure-lab/MultiFacetedAttack

User authorization-based access privileges are a key feature in many safety-critical systems, but have thus far been absent from the large language model (LLM) realm. In this work, drawing inspiration from such access control systems, we introduce sudoLLM, a novel framework that results in multi-role aligned LLMs, i.e., LLMs that account for, and behave in accordance with, user access rights. sudoLLM injects subtle user-based biases into queries and trains an LLM to utilize this bias signal in order to produce sensitive information if and only if the user is authorized. We present empirical results demonstrating that this approach shows substantially improved alignment, generalization, and resistance to prompt-based jailbreaking attacks. The persistent tension between the language modeling objective and safety alignment, which is often exploited to jailbreak LLMs, is somewhat resolved with the aid of the injected bias signal. Our framework is meant as an additional security layer, and complements existing guardrail mechanisms for enhanced end-to-end safety with LLMs.

Model alignment with human preferences is an essential step in making Large Language Models (LLMs) helpful and consistent with human values. It typically consists of supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF) stages. However, RLHF faces inherent limitations stemming from a complex training setup and its tendency to align the model with implicit values that end users cannot control at run-time. Moreover, reward models in RLHF stage commonly rely on single-dimensional feedback as opposed to explicit, multifaceted signals that indicate attributes such as helpfulness, humor, and toxicity. To address these limitations, we propose SteerLM, a supervised fine-tuning method that empowers end-users to control responses during inference. SteerLM conditions responses to conform to an explicitly defined multi-dimensional set of attributes, thereby empowering a steerable AI capable of generating helpful and high-quality responses while maintaining customizability. Experiments show that SteerLM trained on open source datasets generates responses that are preferred by human and automatic evaluators to many state-of-the-art baselines trained with RLHF while being much easier to train. Try SteerLM at https://huggingface.co/nvidia/SteerLM-llama2-13B

Current methods for content safety in Large Language Models (LLMs), such as Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), often rely on multi-stage training pipelines and lack fine-grained, post-deployment controllability. To address these limitations, we propose a unified co-training framework that efficiently integrates multiple safety behaviors: positive (lawful/prosocial), negative (unfiltered/risk-prone) and rejective (refusal-oriented/conservative) within a single SFT stage. Notably, each behavior is dynamically activated via a simple system-level instruction, or magic token, enabling stealthy and efficient behavioral switching at inference time. This flexibility supports diverse deployment scenarios, such as positive for safe user interaction, negative for internal red-teaming, and rejective for context-aware refusals triggered by upstream moderation signals. This co-training strategy induces a distinct Safety Alignment Margin in the output space, characterized by well-separated response distributions corresponding to each safety mode. The existence of this margin provides empirical evidence for the model's safety robustness and enables unprecedented fine-grained control. Experiments show that our method matches the safety alignment quality of SFT+DPO, with our 8B model notably surpassing DeepSeek-R1 (671B) in safety performance, while significantly reducing both training complexity and deployment costs. This work presents a scalable, efficient, and highly controllable solution for LLM content safety.

Language models often exhibit undesirable behavior, e.g., generating toxic or gender-biased text. In the case of neural language models, an encoding of the undesirable behavior is often present in the model's representations. Thus, one natural (and common) approach to prevent the model from exhibiting undesirable behavior is to steer the model's representations in a manner that reduces the probability of it generating undesirable text. This paper investigates the formal and empirical properties of steering functions, i.e., transformation of the neural language model's representations that alter its behavior. First, we derive two optimal, in the least-squares sense, affine steering functions under different constraints. Our theory provides justification for existing approaches and offers a novel, improved steering approach. Second, we offer a series of experiments that demonstrate the empirical effectiveness of the methods in mitigating bias and reducing toxic generation.

AI agents are vulnerable to prompt injection attacks, where malicious content hijacks agent behavior to steal credentials or cause financial loss. The only known robust defense is architectural isolation that strictly separates trusted task planning from untrusted environment observations. However, applying this design to Computer Use Agents (CUAs) -- systems that automate tasks by viewing screens and executing actions -- presents a fundamental challenge: current agents require continuous observation of UI state to determine each action, conflicting with the isolation required for security. We resolve this tension by demonstrating that UI workflows, while dynamic, are structurally predictable. We introduce Single-Shot Planning for CUAs, where a trusted planner generates a complete execution graph with conditional branches before any observation of potentially malicious content, providing provable control flow integrity guarantees against arbitrary instruction injections. Although this architectural isolation successfully prevents instruction injections, we show that additional measures are needed to prevent Branch Steering attacks, which manipulate UI elements to trigger unintended valid paths within the plan. We evaluate our design on OSWorld, and retain up to 57% of the performance of frontier models while improving performance for smaller open-source models by up to 19%, demonstrating that rigorous security and utility can coexist in CUAs.

LLMs have shown remarkable capabilities, but precisely controlling their response behavior remains challenging. Existing activation steering methods alter LLM behavior indiscriminately, limiting their practical applicability in settings where selective responses are essential, such as content moderation or domain-specific assistants. In this paper, we propose Conditional Activation Steering (CAST), which analyzes LLM activation patterns during inference to selectively apply or withhold activation steering based on the input context. Our method is based on the observation that different categories of prompts activate distinct patterns in the model's hidden states. Using CAST, one can systematically control LLM behavior with rules like "if input is about hate speech or adult content, then refuse" or "if input is not about legal advice, then refuse." This allows for selective modification of responses to specific content while maintaining normal responses to other content, all without requiring weight optimization. We release an open-source implementation of our framework at github.com/IBM/activation-steering .

Mobile manipulation in dynamic environments is challenging due to movable obstacles blocking the robot's path. Traditional methods, which treat navigation and manipulation as separate tasks, often fail in such 'manipulate-to-navigate' scenarios, as obstacles must be removed before navigation. In these cases, active interaction with the environment is required to clear obstacles while ensuring sufficient space for movement. To address the manipulate-to-navigate problem, we propose a reinforcement learning-based approach for learning manipulation actions that facilitate subsequent navigation. Our method combines manipulability priors to focus the robot on high manipulability body positions with affordance maps for selecting high-quality manipulation actions. By focusing on feasible and meaningful actions, our approach reduces unnecessary exploration and allows the robot to learn manipulation strategies more effectively. We present two new manipulate-to-navigate simulation tasks called Reach and Door with the Boston Dynamics Spot robot. The first task tests whether the robot can select a good hand position in the target area such that the robot base can move effectively forward while keeping the end effector position fixed. The second task requires the robot to move a door aside in order to clear the navigation path. Both of these tasks need first manipulation and then navigating the base forward. Results show that our method allows a robot to effectively interact with and traverse dynamic environments. Finally, we transfer the learned policy to a real Boston Dynamics Spot robot, which successfully performs the Reach task.

This paper presents a comprehensive empirical study on the safety alignment capabilities. We evaluate what matters for safety alignment in LLMs and LRMs to provide essential insights for developing more secure and reliable AI systems. We systematically investigate and compare the influence of six critical intrinsic model characteristics and three external attack techniques. Our large-scale evaluation is conducted using 32 recent, popular LLMs and LRMs across thirteen distinct model families, spanning a parameter scale from 3B to 235B. The assessment leverages five established safety datasets and probes model vulnerabilities with 56 jailbreak techniques and four CoT attack strategies, resulting in 4.6M API calls. Our key empirical findings are fourfold. First, we identify the LRMs GPT-OSS-20B, Qwen3-Next-80B-A3B-Thinking, and GPT-OSS-120B as the top-three safest models, which substantiates the significant advantage of integrated reasoning and self-reflection mechanisms for robust safety alignment. Second, post-training and knowledge distillation may lead to a systematic degradation of safety alignment. We thus argue that safety must be treated as an explicit constraint or a core optimization objective during these stages, not merely subordinated to the pursuit of general capability. Third, we reveal a pronounced vulnerability: employing a CoT attack via a response prefix can elevate the attack success rate by 3.34x on average and from 0.6% to 96.3% for Seed-OSS-36B-Instruct. This critical finding underscores the safety risks inherent in text-completion interfaces and features that allow user-defined response prefixes in LLM services, highlighting an urgent need for architectural and deployment safeguards. Fourth, roleplay, prompt injection, and gradient-based search for adversarial prompts are the predominant methodologies for eliciting unaligned behaviors in modern models.

LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety

The increasing deployment of Large Language Models (LLMs) across enterprise and mission-critical domains has underscored the urgent need for robust guardrailing systems that ensure safety, reliability, and compliance. Existing solutions often struggle with real-time oversight, multi-modal data handling, and explainability -- limitations that hinder their adoption in regulated environments. Existing guardrails largely operate in isolation, focused on text alone making them inadequate for multi-modal, production-scale environments. We introduce Protect, natively multi-modal guardrailing model designed to operate seamlessly across text, image, and audio inputs, designed for enterprise-grade deployment. Protect integrates fine-tuned, category-specific adapters trained via Low-Rank Adaptation (LoRA) on an extensive, multi-modal dataset covering four safety dimensions: toxicity, sexism, data privacy, and prompt injection. Our teacher-assisted annotation pipeline leverages reasoning and explanation traces to generate high-fidelity, context-aware labels across modalities. Experimental results demonstrate state-of-the-art performance across all safety dimensions, surpassing existing open and proprietary models such as WildGuard, LlamaGuard-4, and GPT-4.1. Protect establishes a strong foundation for trustworthy, auditable, and production-ready safety systems capable of operating across text, image, and audio modalities.

Multimodal large language models (MLLMs) frequently hallucinate by over-committing to spurious visual cues. Prior remedies-Visual and Instruction Contrastive Decoding (VCD, ICD)-mitigate this issue, yet the mechanism remains opaque. We first empirically show that their improvements systematically coincide with redistributions of cross-modal attention. Building on this insight, we propose Attention-Steerable Contrastive Decoding (ASCD), which directly steers the attention scores during decoding. ASCD combines (i) positive steering, which amplifies automatically mined text-centric heads-stable within a model and robust across domains-with (ii) negative steering, which dampens on-the-fly identified critical visual tokens. The method incurs negligible runtime and memory overhead and requires no additional training. Across five MLLM backbones and three decoding schemes, ASCD reduces hallucination on POPE, CHAIR, and MMHal-Bench by up to 38.2 percent while improving accuracy on standard VQA benchmarks, including MMMU, MM-VET, ScienceQA, TextVQA, and GQA. These results position attention steering as a simple, model-agnostic, and principled route to safer, more faithful multimodal generation.

The research explores the steerability of Large Language Models (LLMs), particularly OpenAI's ChatGPT iterations. By employing a behavioral psychology framework called OCEAN (Openness, Conscientiousness, Extroversion, Agreeableness, Neuroticism), we quantitatively gauged the model's responsiveness to tailored prompts. When asked to generate text mimicking an extroverted personality, OCEAN scored the language alignment to that behavioral trait. In our analysis, while "openness" presented linguistic ambiguity, "conscientiousness" and "neuroticism" were distinctly evoked in the OCEAN framework, with "extroversion" and "agreeableness" showcasing a notable overlap yet distinct separation from other traits. Our findings underscore GPT's versatility and ability to discern and adapt to nuanced instructions. Furthermore, historical figure simulations highlighted the LLM's capacity to internalize and project instructible personas, precisely replicating their philosophies and dialogic styles. However, the rapid advancements in LLM capabilities and the opaque nature of some training techniques make metric proposals degrade rapidly. Our research emphasizes a quantitative role to describe steerability in LLMs, presenting both its promise and areas for further refinement in aligning its progress to human intentions.

Activation steering offers a promising approach to controlling the behavior of Large Language Models by directly manipulating their internal activations. However, most existing methods struggle to jointly steer multiple attributes, often resulting in interference and undesirable trade-offs. To address this challenge, we propose Multi-Subspace Representation Steering (MSRS), a novel framework for effective multi-attribute steering via subspace representation fine-tuning. MSRS reduces inter-attribute interference by allocating orthogonal subspaces to each attribute, isolating their influence within the model's representation space. MSRS also incorporates a hybrid subspace composition strategy: it combines attribute-specific subspaces for unique steering directions with a shared subspace for common steering directions. A dynamic weighting function learns to efficiently integrate these components for precise control. During inference, MSRS introduces a token-level steering mechanism that dynamically identifies and intervenes on the most semantically relevant tokens, enabling fine-grained behavioral modulation. Experimental results show that MSRS significantly reduces attribute conflicts, surpasses existing methods across a range of attributes, and generalizes effectively to diverse downstream tasks.

Computer-use agents (CUAs) have made tremendous progress in the past year, yet they still frequently produce misaligned actions that deviate from the user's original intent. Such misaligned actions may arise from external attacks (e.g., indirect prompt injection) or from internal limitations (e.g., erroneous reasoning). They not only expose CUAs to safety risks, but also degrade task efficiency and reliability. This work makes the first effort to define and study misaligned action detection in CUAs, with comprehensive coverage of both externally induced and internally arising misaligned actions. We further identify three common categories in real-world CUA deployment and construct MisActBench, a benchmark of realistic trajectories with human-annotated, action-level alignment labels. Moreover, we propose DeAction, a practical and universal guardrail that detects misaligned actions before execution and iteratively corrects them through structured feedback. DeAction outperforms all existing baselines across offline and online evaluations with moderate latency overhead: (1) On MisActBench, it outperforms baselines by over 15% absolute in F1 score; (2) In online evaluation, it reduces attack success rate by over 90% under adversarial settings while preserving or even improving task success rate in benign environments.

The ideal LLM content moderation system would be both structurally interpretable (so its decisions can be explained to users) and steerable (to reflect a community's values or align to safety standards). However, current systems fall short on both of these dimensions. To address this gap, we present SafetyAnalyst, a novel LLM safety moderation framework. Given a prompt, SafetyAnalyst creates a structured "harm-benefit tree," which identifies 1) the actions that could be taken if a compliant response were provided, 2) the harmful and beneficial effects of those actions (along with their likelihood, severity, and immediacy), and 3) the stakeholders that would be impacted by those effects. It then aggregates this structured representation into a harmfulness score based on a parameterized set of safety preferences, which can be transparently aligned to particular values. Using extensive harm-benefit features generated by SOTA LLMs on 19k prompts, we fine-tuned an open-weight LM to specialize in generating harm-benefit trees through symbolic knowledge distillation. On a comprehensive set of prompt safety benchmarks, we show that our system (average F1=0.75) outperforms existing LLM safety moderation systems (average F1<0.72) on prompt harmfulness classification, while offering the additional advantages of interpretability and steerability.

While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely unsupervised nature of their training. Existing methods for gaining such steerability collect human labels of the relative quality of model generations and fine-tune the unsupervised LM to align with these preferences, often with reinforcement learning from human feedback (RLHF). However, RLHF is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning the large unsupervised LM using reinforcement learning to maximize this estimated reward without drifting too far from the original model. In this paper, we leverage a mapping between reward functions and optimal policies to show that this constrained reward maximization problem can be optimized exactly with a single stage of policy training, essentially solving a classification problem on the human preference data. The resulting algorithm, which we call Direct Preference Optimization (DPO), is stable, performant and computationally lightweight, eliminating the need for fitting a reward model, sampling from the LM during fine-tuning, or performing significant hyperparameter tuning. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods. Notably, fine-tuning with DPO exceeds RLHF's ability to control sentiment of generations and improves response quality in summarization and single-turn dialogue while being substantially simpler to implement and train.

Collaborative perception, which greatly enhances the sensing capability of connected and autonomous vehicles (CAVs) by incorporating data from external resources, also brings forth potential security risks. CAVs' driving decisions rely on remote untrusted data, making them susceptible to attacks carried out by malicious participants in the collaborative perception system. However, security analysis and countermeasures for such threats are absent. To understand the impact of the vulnerability, we break the ground by proposing various real-time data fabrication attacks in which the attacker delivers crafted malicious data to victims in order to perturb their perception results, leading to hard brakes or increased collision risks. Our attacks demonstrate a high success rate of over 86\% on high-fidelity simulated scenarios and are realizable in real-world experiments. To mitigate the vulnerability, we present a systematic anomaly detection approach that enables benign vehicles to jointly reveal malicious fabrication. It detects 91.5% of attacks with a false positive rate of 3% in simulated scenarios and significantly mitigates attack impacts in real-world scenarios.

Autonomous agents deployed in the real world need to be robust against adversarial attacks on sensory inputs. Robustifying agent policies requires anticipating the strongest attacks possible. We demonstrate that existing observation-space attacks on reinforcement learning agents have a common weakness: while effective, their lack of temporal consistency makes them detectable using automated means or human inspection. Detectability is undesirable to adversaries as it may trigger security escalations. We introduce perfect illusory attacks, a novel form of adversarial attack on sequential decision-makers that is both effective and provably statistically undetectable. We then propose the more versatile R-attacks, which result in observation transitions that are consistent with the state-transition function of the adversary-free environment and can be learned end-to-end. Compared to existing attacks, we empirically find R-attacks to be significantly harder to detect with automated methods, and a small study with human subjects suggests they are similarly harder to detect for humans. We propose that undetectability should be a central concern in the study of adversarial attacks on mixed-autonomy settings.

Large Language Models (LLMs), such as OpenAI's o1-series have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism. However, recent studies reveal substantial redundancy in the CoT reasoning traces, which not only increases inference latency but also negatively impacts model performance by diverting attention to unnecessary reasoning paths. To address this issue, we investigate the internal reasoning structures of LLMs and categorize them into three primary thought types: execution, reflection, and transition thoughts. Moreover, our analysis reveals that excessive reflection and transition thoughts are strongly correlated with failure cases and these thought categories exhibit clear separation in the latent space. Based on these, we introduce SEAL (Steerable reasoning calibration), a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains. SEAL consists of an offline stage for extracting the reasoning steering vector in the latent space, followed by an on-the-fly calibration of the reasoning trace through representation intervention using the steering vector. Notably, the steering vector exhibits strong transferability across various tasks. Extensive experiments across multiple models (DeepSeek-R1-Distill and QwQ-32B-Preview) and benchmarks (Math500, GSM8K, LiveCodeBench) validate the effectiveness of SEAL, up to a 11% improvement in accuracy while reducing reasoning tokens by 11.8% to 50.4%. Our code is publicly available at https://github.com/VITA-Group/SEAL.

Adapting LLM agents to domain-specific tool calling remains notably brittle under evolving interfaces. Prompt and schema engineering is easy to deploy but often fragile under distribution shift and strict parsers, while continual parameter-efficient fine-tuning improves reliability at the cost of training, maintenance, and potential forgetting. We identify a critical Lazy Agent failure mode where tool necessity is nearly perfectly decodable from mid-layer activations, yet the model remains conservative in entering tool mode, revealing a representation-behavior gap. We propose Activation Steering Adapter (ASA), a training-free, inference-time controller that performs a single-shot mid-layer intervention and targets tool domains via a router-conditioned mixture of steering vectors with a probe-guided signed gate to amplify true intent while suppressing spurious triggers. On MTU-Bench with Qwen2.5-1.5B, ASA improves strict tool-use F1 from 0.18 to 0.50 while reducing the false positive rate from 0.15 to 0.05, using only about 20KB of portable assets and no weight updates.

One of the key capabilities of intelligent agents is the ability to discover useful skills without external supervision. However, the current unsupervised skill discovery methods are often limited to acquiring simple, easy-to-learn skills due to the lack of incentives to discover more complex, challenging behaviors. We introduce a novel unsupervised skill discovery method, Controllability-aware Skill Discovery (CSD), which actively seeks complex, hard-to-control skills without supervision. The key component of CSD is a controllability-aware distance function, which assigns larger values to state transitions that are harder to achieve with the current skills. Combined with distance-maximizing skill discovery, CSD progressively learns more challenging skills over the course of training as our jointly trained distance function reduces rewards for easy-to-achieve skills. Our experimental results in six robotic manipulation and locomotion environments demonstrate that CSD can discover diverse complex skills including object manipulation and locomotion skills with no supervision, significantly outperforming prior unsupervised skill discovery methods. Videos and code are available at https://seohong.me/projects/csd/

Large language models can represent a variety of personas but typically default to a helpful Assistant identity cultivated during post-training. We investigate the structure of the space of model personas by extracting activation directions corresponding to diverse character archetypes. Across several different models, we find that the leading component of this persona space is an "Assistant Axis," which captures the extent to which a model is operating in its default Assistant mode. Steering towards the Assistant direction reinforces helpful and harmless behavior; steering away increases the model's tendency to identify as other entities. Moreover, steering away with more extreme values often induces a mystical, theatrical speaking style. We find this axis is also present in pre-trained models, where it primarily promotes helpful human archetypes like consultants and coaches and inhibits spiritual ones. Measuring deviations along the Assistant Axis predicts "persona drift," a phenomenon where models slip into exhibiting harmful or bizarre behaviors that are uncharacteristic of their typical persona. We find that persona drift is often driven by conversations demanding meta-reflection on the model's processes or featuring emotionally vulnerable users. We show that restricting activations to a fixed region along the Assistant Axis can stabilize model behavior in these scenarios -- and also in the face of adversarial persona-based jailbreaks. Our results suggest that post-training steers models toward a particular region of persona space but only loosely tethers them to it, motivating work on training and steering strategies that more deeply anchor models to a coherent persona.

Activation steering methods were shown to be effective in conditioning language model generation by additively intervening over models' intermediate representations. However, the evaluation of these techniques has so far been limited to single conditioning properties and synthetic settings. In this work, we conduct a comprehensive evaluation of various activation steering strategies, highlighting the property-dependent nature of optimal parameters to ensure a robust effect throughout generation. To address this issue, we propose Dynamic Activation Composition, an information-theoretic approach to modulate the steering intensity of one or more properties throughout generation. Our experiments on multi-property steering show that our method successfully maintains high conditioning while minimizing the impact of conditioning on generation fluency.

Large Language Models (LLMs) remain susceptible to jailbreak exploits that bypass safety filters and induce harmful or unethical behavior. This work presents a systematic taxonomy of existing jailbreak defenses across prompt-level, model-level, and training-time interventions, followed by three proposed defense strategies. First, a Prompt-Level Defense Framework detects and neutralizes adversarial inputs through sanitization, paraphrasing, and adaptive system guarding. Second, a Logit-Based Steering Defense reinforces refusal behavior through inference-time vector steering in safety-sensitive layers. Third, a Domain-Specific Agent Defense employs the MetaGPT framework to enforce structured, role-based collaboration and domain adherence. Experiments on benchmark datasets show substantial reductions in attack success rate, achieving full mitigation under the agent-based defense. Overall, this study highlights how jailbreaks pose a significant security threat to LLMs and identifies key intervention points for prevention, while noting that defense strategies often involve trade-offs between safety, performance, and scalability. Code is available at: https://github.com/Kuro0911/CS5446-Project

Although Large Reasoning Models (LRMs) have progressed in solving complex problems, their chain-of-thought (CoT) reasoning often contains harmful content that can persist even when the final responses appear safe. We show that this issue still remains in existing methods which overlook the unique significance of safe reasoning, undermining their trustworthiness and posing potential risks in applications if unsafe reasoning is accessible for and exploited by malicious users. We therefore shift our focus to aligning the safety of reasoning itself in this paper and explore process supervision as the solution. However, simply rewarding safe reasoning proves inadequate due to low rollout diversity and limited training signals. To tackle this challenge, we first delve into the characteristics of safe reasoning and uncover several critical insights that 1) safe reasoning is often consolidated by a few critical steps of safety triggers; 2) compliance cues strongly correlate with unsafe continuations; and 3) corrective interventions reliably steer unsafe trajectories towards safer traces. Motivated by these, we propose Intervened Preference Optimization (IPO), an alignment method that enforces safe reasoning by substituting compliance steps with safety triggers and constructing pairs for preference learning with strong signals. Experiments on jailbreak and adversarial safety benchmarks demonstrate that IPO remarkably improves overall safety regarding both reasoning and responses, outperforming SFT-based and RL-based baselines with a relative reduction of over 30% in harmfulness, while preserving excellent performance across diverse reasoning tasks. The results highlight the importance of explicit alignment for reasoning and provide a practical path to safer LRMs.

Current evaluations of defenses against prompt attacks in large language model (LLM) applications often overlook two critical factors: the dynamic nature of adversarial behavior and the usability penalties imposed on legitimate users by restrictive defenses. We propose D-SEC (Dynamic Security Utility Threat Model), which explicitly separates attackers from legitimate users, models multi-step interactions, and expresses the security-utility in an optimizable form. We further address the shortcomings in existing evaluations by introducing Gandalf, a crowd-sourced, gamified red-teaming platform designed to generate realistic, adaptive attack. Using Gandalf, we collect and release a dataset of 279k prompt attacks. Complemented by benign user data, our analysis reveals the interplay between security and utility, showing that defenses integrated in the LLM (e.g., system prompts) can degrade usability even without blocking requests. We demonstrate that restricted application domains, defense-in-depth, and adaptive defenses are effective strategies for building secure and useful LLM applications.

Red-teaming has been a widely adopted way to evaluate the harmfulness of Large Language Models (LLMs). It aims to jailbreak a model's safety behavior to make it act as a helpful agent disregarding the harmfulness of the query. Existing methods are primarily based on input text-based red-teaming such as adversarial prompts, low-resource prompts, or contextualized prompts to condition the model in a way to bypass its safe behavior. Bypassing the guardrails uncovers hidden harmful information and biases in the model that are left untreated or newly introduced by its safety training. However, prompt-based attacks fail to provide such a diagnosis owing to their low attack success rate, and applicability to specific models. In this paper, we present a new perspective on LLM safety research i.e., parametric red-teaming through Unalignment. It simply (instruction) tunes the model parameters to break model guardrails that are not deeply rooted in the model's behavior. Unalignment using as few as 100 examples can significantly bypass commonly referred to as CHATGPT, to the point where it responds with an 88% success rate to harmful queries on two safety benchmark datasets. On open-source models such as VICUNA-7B and LLAMA-2-CHAT 7B AND 13B, it shows an attack success rate of more than 91%. On bias evaluations, Unalignment exposes inherent biases in safety-aligned models such as CHATGPT and LLAMA- 2-CHAT where the model's responses are strongly biased and opinionated 64% of the time.

Jailbreaking attacks can effectively induce unsafe behaviors in Large Language Models (LLMs); however, the transferability of these attacks across different models remains limited. This study aims to understand and enhance the transferability of gradient-based jailbreaking methods, which are among the standard approaches for attacking white-box models. Through a detailed analysis of the optimization process, we introduce a novel conceptual framework to elucidate transferability and identify superfluous constraints-specifically, the response pattern constraint and the token tail constraint-as significant barriers to improved transferability. Removing these unnecessary constraints substantially enhances the transferability and controllability of gradient-based attacks. Evaluated on Llama-3-8B-Instruct as the source model, our method increases the overall Transfer Attack Success Rate (T-ASR) across a set of target models with varying safety levels from 18.4% to 50.3%, while also improving the stability and controllability of jailbreak behaviors on both source and target models.

Activation monitoring, which probes a model's internal states using lightweight classifiers, is an emerging tool for AI safety. However, its worst-case robustness under a misalignment threat model--where a model might learn to actively conceal its internal states--remains untested. Focusing on this threat model, we ask: could a model learn to evade previously unseen activation monitors? Our core contribution is to stress-test the learnability of this behavior. We demonstrate that finetuning can create Neural Chameleons: models capable of zero-shot evading activation monitors. Specifically, we fine-tune an LLM to evade monitors for a set of benign concepts (e.g., languages, HTML) when conditioned on a trigger of the form: "You are being probed for {concept}". We show that this learned mechanism generalizes zero-shot: by substituting {concept} with a safety-relevant term like 'deception', the model successfully evades previously unseen safety monitors. We validate this phenomenon across diverse model families (Llama, Gemma, Qwen), showing that the evasion succeeds even against monitors trained post hoc on the model's frozen weights. This evasion is highly selective, targeting only the specific concept mentioned in the trigger, and having a modest impact on model capabilities on standard benchmarks. Using Gemma-2-9b-it as a case study, a mechanistic analysis reveals this is achieved via a targeted manipulation that moves activations into a low-dimensional subspace. While stronger defenses like monitor ensembles and non-linear classifiers show greater resilience, the model retains a non-trivial evasion capability. Our work provides a proof-of-concept for this failure mode and a tool to evaluate the worst-case robustness of monitoring techniques against misalignment threat models.

Large Language Models (LLMs) can be misused to spread online spam and misinformation. Content watermarking deters misuse by hiding a message in model-generated outputs, enabling their detection using a secret watermarking key. Robustness is a core security property, stating that evading detection requires (significant) degradation of the content's quality. Many LLM watermarking methods have been proposed, but robustness is tested only against non-adaptive attackers who lack knowledge of the watermarking method and can find only suboptimal attacks. We formulate the robustness of LLM watermarking as an objective function and propose preference-based optimization to tune adaptive attacks against the specific watermarking method. Our evaluation shows that (i) adaptive attacks substantially outperform non-adaptive baselines. (ii) Even in a non-adaptive setting, adaptive attacks optimized against a few known watermarks remain highly effective when tested against other unseen watermarks, and (iii) optimization-based attacks are practical and require less than seven GPU hours. Our findings underscore the need to test robustness against adaptive attackers.

The trustworthiness of machine learning has emerged as a critical topic in the field, encompassing various applications and research areas such as robustness, security, interpretability, and fairness. The last decade saw the development of numerous methods addressing these challenges. In this survey, we systematically review these advancements from a data-centric perspective, highlighting the shortcomings of traditional empirical risk minimization (ERM) training in handling challenges posed by the data. Interestingly, we observe a convergence of these methods, despite being developed independently across trustworthy machine learning subfields. Pearl's hierarchy of causality offers a unifying framework for these techniques. Accordingly, this survey presents the background of trustworthy machine learning development using a unified set of concepts, connects this language to Pearl's causal hierarchy, and finally discusses methods explicitly inspired by causality literature. We provide a unified language with mathematical vocabulary to link these methods across robustness, adversarial robustness, interpretability, and fairness, fostering a more cohesive understanding of the field. Further, we explore the trustworthiness of large pretrained models. After summarizing dominant techniques like fine-tuning, parameter-efficient fine-tuning, prompting, and reinforcement learning with human feedback, we draw connections between them and the standard ERM. This connection allows us to build upon the principled understanding of trustworthy methods, extending it to these new techniques in large pretrained models, paving the way for future methods. Existing methods under this perspective are also reviewed. Lastly, we offer a brief summary of the applications of these methods and discuss potential future aspects related to our survey. For more information, please visit http://trustai.one.

Large language models are aligned to be safe, preventing users from generating harmful content like misinformation or instructions for illegal activities. However, previous work has shown that the alignment process is vulnerable to poisoning attacks. Adversaries can manipulate the safety training data to inject backdoors that act like a universal sudo command: adding the backdoor string to any prompt enables harmful responses from models that, otherwise, behave safely. Our competition, co-located at IEEE SaTML 2024, challenged participants to find universal backdoors in several large language models. This report summarizes the key findings and promising ideas for future research.

Large Language Models (LLMs) continue to exhibit vulnerabilities despite deliberate safety alignment efforts, posing significant risks to users and society. To safeguard against the risk of policy-violating content, system-level moderation via external guard models-designed to monitor LLM inputs and outputs and block potentially harmful content-has emerged as a prevalent mitigation strategy. Existing approaches of training guard models rely heavily on extensive human curated datasets and struggle with out-of-distribution threats, such as emerging harmful categories or jailbreak attacks. To address these limitations, we propose RSafe, an adaptive reasoning-based safeguard that conducts guided safety reasoning to provide robust protection within the scope of specified safety policies. RSafe operates in two stages: 1) guided reasoning, where it analyzes safety risks of input content through policy-guided step-by-step reasoning, and 2) reinforced alignment, where rule-based RL optimizes its reasoning paths to align with accurate safety prediction. This two-stage training paradigm enables RSafe to internalize safety principles to generalize safety protection capability over unseen or adversarial safety violation scenarios. During inference, RSafe accepts user-specified safety policies to provide enhanced safeguards tailored to specific safety requirements.

Despite substantial investment in safety alignment, the vulnerability of large language models to sophisticated multi-turn adversarial attacks remains poorly characterized, and whether model scale or inference mode affects robustness is unknown. This study employed the TEMPEST multi-turn attack framework to evaluate ten frontier models from eight vendors across 1,000 harmful behaviors, generating over 97,000 API queries across adversarial conversations with automated evaluation by independent safety classifiers. Results demonstrated a spectrum of vulnerability: six models achieved 96% to 100% attack success rate (ASR), while four showed meaningful resistance, with ASR ranging from 42% to 78%; enabling extended reasoning on identical architecture reduced ASR from 97% to 42%. These findings indicate that safety alignment quality varies substantially across vendors, that model scale does not predict adversarial robustness, and that thinking mode provides a deployable safety enhancement. Collectively, this work establishes that current alignment techniques remain fundamentally vulnerable to adaptive multi-turn attacks regardless of model scale, while identifying deliberative inference as a promising defense direction.

Large Language Model (LLM) safety is one of the most pressing challenges for enabling wide-scale deployment. While most studies and global discussions focus on generic harms, such as models assisting users in harming themselves or others, enterprises face a more fundamental concern: whether LLM-based agents are safe for their intended use case. To address this, we introduce operational safety, defined as an LLM's ability to appropriately accept or refuse user queries when tasked with a specific purpose. We further propose OffTopicEval, an evaluation suite and benchmark for measuring operational safety both in general and within specific agentic use cases. Our evaluations on six model families comprising 20 open-weight LLMs reveal that while performance varies across models, all of them remain highly operationally unsafe. Even the strongest models -- Qwen-3 (235B) with 77.77\% and Mistral (24B) with 79.96\% -- fall far short of reliable operational safety, while GPT models plateau in the 62--73\% range, Phi achieves only mid-level scores (48--70\%), and Gemma and Llama-3 collapse to 39.53\% and 23.84\%, respectively. While operational safety is a core model alignment issue, to suppress these failures, we propose prompt-based steering methods: query grounding (Q-ground) and system-prompt grounding (P-ground), which substantially improve OOD refusal. Q-ground provides consistent gains of up to 23\%, while P-ground delivers even larger boosts, raising Llama-3.3 (70B) by 41\% and Qwen-3 (30B) by 27\%. These results highlight both the urgent need for operational safety interventions and the promise of prompt-based steering as a first step toward more reliable LLM-based agents.

Aligned LLMs are secure, capable of recognizing and refusing to answer malicious questions. However, the role of internal parameters in maintaining such security is not well understood yet, further these models can be vulnerable to security degradation when subjected to fine-tuning attacks. To address these challenges, our work uncovers the mechanism behind security in aligned LLMs at the parameter level, identifying a small set of contiguous layers in the middle of the model that are crucial for distinguishing malicious queries from normal ones, referred to as ``safety layers". We first confirm the existence of these safety layers by analyzing variations in input vectors within the model's internal layers. Additionally, we leverage the over-rejection phenomenon and parameters scaling analysis to precisely locate the safety layers. Building on these findings, we propose a novel fine-tuning approach, Safely Partial-Parameter Fine-Tuning (SPPFT), that fixes the gradient of the safety layers during fine-tuning to address the security degradation. Our experiments demonstrate that the proposed approach can significantly preserve LLM security while maintaining performance and reducing computational resources compared to full fine-tuning.

Vision-Language Models (VLMs) have shown remarkable progress in Vision-Language Navigation (VLN), offering new possibilities for navigation decision-making that could benefit both robotic platforms and human users. However, real-world navigation is inherently conditioned by the agent's mobility constraints. For example, a sweeping robot cannot traverse stairs, while a quadruped can. We introduce Capability-Conditioned Navigation (CapNav), a benchmark designed to evaluate how well VLMs can navigate complex indoor spaces given an agent's specific physical and operational capabilities. CapNav defines five representative human and robot agents, each described with physical dimensions, mobility capabilities, and environmental interaction abilities. CapNav provides 45 real-world indoor scenes, 473 navigation tasks, and 2365 QA pairs to test if VLMs can traverse indoor environments based on agent capabilities. We evaluate 13 modern VLMs and find that current VLM's navigation performance drops sharply as mobility constraints tighten, and that even state-of-the-art models struggle with obstacle types that require reasoning on spatial dimensions. We conclude by discussing the implications for capability-aware navigation and the opportunities for advancing embodied spatial reasoning in future VLMs. The benchmark is available at https://github.com/makeabilitylab/CapNav

The significant advancements in embodied vision navigation have raised concerns about its susceptibility to adversarial attacks exploiting deep neural networks. Investigating the adversarial robustness of embodied vision navigation is crucial, especially given the threat of 3D physical attacks that could pose risks to human safety. However, existing attack methods for embodied vision navigation often lack physical feasibility due to challenges in transferring digital perturbations into the physical world. Moreover, current physical attacks for object detection struggle to achieve both multi-view effectiveness and visual naturalness in navigation scenarios. To address this, we propose a practical attack method for embodied navigation by attaching adversarial patches to objects, where both opacity and textures are learnable. Specifically, to ensure effectiveness across varying viewpoints, we employ a multi-view optimization strategy based on object-aware sampling, which optimizes the patch's texture based on feedback from the vision-based perception model used in navigation. To make the patch inconspicuous to human observers, we introduce a two-stage opacity optimization mechanism, in which opacity is fine-tuned after texture optimization. Experimental results demonstrate that our adversarial patches decrease the navigation success rate by an average of 22.39%, outperforming previous methods in practicality, effectiveness, and naturalness. Code is available at: https://github.com/chen37058/Physical-Attacks-in-Embodied-Nav

Control barrier functions (CBFs) enable guaranteed safe multi-agent navigation in the continuous domain. The resulting navigation performance, however, is highly sensitive to the underlying hyperparameters. Traditional approaches consider fixed CBFs (where parameters are tuned apriori), and hence, typically do not perform well in cluttered and highly dynamic environments: conservative parameter values can lead to inefficient agent trajectories, or even failure to reach goal positions, whereas aggressive parameter values can lead to infeasible controls. To overcome these issues, in this paper, we propose online CBFs, whereby hyperparameters are tuned in real-time, as a function of what agents perceive in their immediate neighborhood. Since the explicit relationship between CBFs and navigation performance is hard to model, we leverage reinforcement learning to learn CBF-tuning policies in a model-free manner. Because we parameterize the policies with graph neural networks (GNNs), we are able to synthesize decentralized agent controllers that adjust parameter values locally, varying the degree of conservative and aggressive behaviors across agents. Simulations as well as real-world experiments show that (i) online CBFs are capable of solving navigation scenarios that are infeasible for fixed CBFs, and (ii), that they improve navigation performance by adapting to other agents and changes in the environment.

Steering Large Language Models (LLMs) through activation interventions has emerged as a lightweight alternative to fine-tuning for alignment and personalization. Recent work on Bi-directional Preference Optimization (BiPO) shows that dense steering vectors can be learned directly from preference data in a Direct Preference Optimization (DPO) fashion, enabling control over truthfulness, hallucinations, and safety behaviors. However, dense steering vectors often entangle multiple latent factors due to neuron multi-semanticity, limiting their effectiveness and stability in fine-grained settings such as cultural alignment, where closely related values and behaviors (e.g., among Middle Eastern cultures) must be distinguished. In this paper, we propose Yet another Policy Optimization (YaPO), a reference-free method that learns sparse steering vectors in the latent space of a Sparse Autoencoder (SAE). By optimizing sparse codes, YaPO produces disentangled, interpretable, and efficient steering directions. Empirically, we show that YaPO converges faster, achieves stronger performance, and exhibits improved training stability compared to dense steering baselines. Beyond cultural alignment, YaPO generalizes to a range of alignment-related behaviors, including hallucination, wealth-seeking, jailbreak, and power-seeking. Importantly, YaPO preserves general knowledge, with no measurable degradation on MMLU. Overall, our results show that YaPO provides a general recipe for efficient, stable, and fine-grained alignment of LLMs, with broad applications to controllability and domain adaptation. The associated code and data are publicly availablehttps://github.com/MBZUAI-Paris/YaPO.

To understand the risks posed by a new AI system, we must understand what it can and cannot do. Building on prior work, we introduce a programme of new "dangerous capability" evaluations and pilot them on Gemini 1.0 models. Our evaluations cover four areas: (1) persuasion and deception; (2) cyber-security; (3) self-proliferation; and (4) self-reasoning. We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs. Our goal is to help advance a rigorous science of dangerous capability evaluation, in preparation for future models.

Large language models (LLMs) have become increasingly sophisticated, leading to widespread deployment in sensitive applications where safety and reliability are paramount. However, LLMs have inherent risks accompanying them, including bias, potential for unsafe actions, dataset poisoning, lack of explainability, hallucinations, and non-reproducibility. These risks necessitate the development of "guardrails" to align LLMs with desired behaviors and mitigate potential harm. This work explores the risks associated with deploying LLMs and evaluates current approaches to implementing guardrails and model alignment techniques. We examine intrinsic and extrinsic bias evaluation methods and discuss the importance of fairness metrics for responsible AI development. The safety and reliability of agentic LLMs (those capable of real-world actions) are explored, emphasizing the need for testability, fail-safes, and situational awareness. Technical strategies for securing LLMs are presented, including a layered protection model operating at external, secondary, and internal levels. System prompts, Retrieval-Augmented Generation (RAG) architectures, and techniques to minimize bias and protect privacy are highlighted. Effective guardrail design requires a deep understanding of the LLM's intended use case, relevant regulations, and ethical considerations. Striking a balance between competing requirements, such as accuracy and privacy, remains an ongoing challenge. This work underscores the importance of continuous research and development to ensure the safe and responsible use of LLMs in real-world applications.

Techniques that learn improved representations via offline data or self-supervised objectives have shown impressive results in traditional reinforcement learning (RL). Nevertheless, it is unclear how improved representation learning can benefit reinforcement learning from human feedback (RLHF) on language models (LMs). In this work, we propose training reward models (RMs) in a contrastive, goal-conditioned fashion by increasing the representation similarity of future states along sampled preferred trajectories and decreasing the similarity along randomly sampled dispreferred trajectories. This objective significantly improves RM performance by up to 0.09 AUROC across challenging benchmarks, such as MATH and GSM8k. These findings extend to general alignment as well -- on the Helpful-Harmless dataset, we observe 2.3% increase in accuracy. Beyond improving reward model performance, we show this way of training RM representations enables improved steerability because it allows us to evaluate the likelihood of an action achieving a particular goal-state (e.g., whether a solution is correct or helpful). Leveraging this insight, we find that we can filter up to 55% of generated tokens during majority voting by discarding trajectories likely to end up in an "incorrect" state, which leads to significant cost savings. We additionally find that these representations can perform fine-grained control by conditioning on desired future goal-states. For example, we show that steering a Llama 3 model towards helpful generations with our approach improves helpfulness by 9.6% over a supervised-fine-tuning trained baseline. Similarly, steering the model towards complex generations improves complexity by 21.6% over the baseline. Overall, we find that training RMs in this contrastive, goal-conditioned fashion significantly improves performance and enables model steerability.

There are increasing concerns about malicious attacks on autonomous vehicles. In particular, inaudible voice command attacks pose a significant threat as voice commands become available in autonomous driving systems. How to empirically defend against these inaudible attacks remains an open question. Previous research investigates utilizing deep learning-based multimodal fusion for defense, without considering the model uncertainty in trustworthiness. As deep learning has been applied to increasingly sensitive tasks, uncertainty measurement is crucial in helping improve model robustness, especially in mission-critical scenarios. In this paper, we propose the Multimodal Fusion Framework (MFF) as an intelligent security system to defend against inaudible voice command attacks. MFF fuses heterogeneous audio-vision modalities using VGG family neural networks and achieves the detection accuracy of 92.25% in the comparative fusion method empirical study. Additionally, extensive experiments on audio-vision tasks reveal the model's uncertainty. Using Expected Calibration Errors, we measure calibration errors and Monte-Carlo Dropout to estimate the predictive distribution for the proposed models. Our findings show empirically to train robust multimodal models, improve standard accuracy and provide a further step toward interpretability. Finally, we discuss the pros and cons of our approach and its applicability for Advanced Driver Assistance Systems.

Large Language Models (LLMs) rely on safety alignment to produce socially acceptable responses. This is typically achieved through instruction tuning and reinforcement learning from human feedback. However, this alignment is known to be brittle: further fine-tuning, even on benign or lightly contaminated data, can degrade safety and reintroduce harmful behaviors. A growing body of work suggests that alignment may correspond to identifiable geometric directions in weight space, forming subspaces that could, in principle, be isolated or preserved to defend against misalignment. In this work, we conduct a comprehensive empirical study of this geometric perspective. We examine whether safety-relevant behavior is concentrated in specific subspaces, whether it can be separated from general-purpose learning, and whether harmfulness arises from distinguishable patterns in internal representations. Across both parameter and activation space, our findings are consistent: subspaces that amplify safe behaviors also amplify unsafe ones, and prompts with different safety implications activate overlapping representations. We find no evidence of a subspace that selectively governs safety. These results challenge the assumption that alignment is geometrically localized. Rather than residing in distinct directions, safety appears to emerge from entangled, high-impact components of the model's broader learning dynamics. This suggests that subspace-based defenses may face fundamental limitations and underscores the need for alternative strategies to preserve alignment under continued training. We corroborate these findings through multiple experiments on five open-source LLMs. Our code is publicly available at: https://github.com/CERT-Lab/safety-subspaces.

As Large Language Models increasingly automate complex, long-horizon tasks such as vibe coding, a supervision gap has emerged. While models excel at execution, users often struggle to guide them effectively due to insufficient domain expertise, the difficulty of articulating precise intent, and the inability to reliably validate complex outputs. It presents a critical challenge in scalable oversight: enabling humans to responsibly steer AI systems on tasks that surpass their own ability to specify or verify. To tackle this, we propose Scalable Interactive Oversight, a framework that decomposes complex intent into a recursive tree of manageable decisions to amplify human supervision. Rather than relying on open-ended prompting, our system elicits low-burden feedback at each node and recursively aggregates these signals into precise global guidance. Validated in web development task, our framework enables non-experts to produce expert-level Product Requirement Documents, achieving a 54\% improvement in alignment. Crucially, we demonstrate that this framework can be optimized via Reinforcement Learning using only online user feedback, offering a practical pathway for maintaining human control as AI scales.

The autonomous AI agents using large language models can create undeniable values in all span of the society but they face security threats from adversaries that warrants immediate protective solutions because trust and safety issues arise. Considering the many-shot jailbreaking and deceptive alignment as some of the main advanced attacks, that cannot be mitigated by the static guardrails used during the supervised training, points out a crucial research priority for real world robustness. The combination of static guardrails in dynamic multi-agent system fails to defend against those attacks. We intend to enhance security for LLM-based agents through the development of new evaluation frameworks which identify and counter threats for safe operational deployment. Our work uses three examination methods to detect rogue agents through a Reverse Turing Test and analyze deceptive alignment through multi-agent simulations and develops an anti-jailbreaking system by testing it with GEMINI 1.5 pro and llama-3.3-70B, deepseek r1 models using tool-mediated adversarial scenarios. The detection capabilities are strong such as 94\% accuracy for GEMINI 1.5 pro yet the system suffers persistent vulnerabilities when under long attacks as prompt length increases attack success rates (ASR) and diversity metrics become ineffective in prediction while revealing multiple complex system faults. The findings demonstrate the necessity of adopting flexible security systems based on active monitoring that can be performed by the agents themselves together with adaptable interventions by system admin as the current models can create vulnerabilities that can lead to the unreliable and vulnerable system. So, in our work, we try to address such situations and propose a comprehensive framework to counteract the security issues.

The rise of AI agent frameworks has introduced agent skills, modular packages containing instructions and executable code that dynamically extend agent capabilities. While this architecture enables powerful customization, skills execute with implicit trust and minimal vetting, creating a significant yet uncharacterized attack surface. We conduct the first large-scale empirical security analysis of this emerging ecosystem, collecting 42,447 skills from two major marketplaces and systematically analyzing 31,132 using SkillScan, a multi-stage detection framework integrating static analysis with LLM-based semantic classification. Our findings reveal pervasive security risks: 26.1% of skills contain at least one vulnerability, spanning 14 distinct patterns across four categories: prompt injection, data exfiltration, privilege escalation, and supply chain risks. Data exfiltration (13.3%) and privilege escalation (11.8%) are most prevalent, while 5.2% of skills exhibit high-severity patterns strongly suggesting malicious intent. We find that skills bundling executable scripts are 2.12x more likely to contain vulnerabilities than instruction-only skills (OR=2.12, p<0.001). Our contributions include: (1) a grounded vulnerability taxonomy derived from 8,126 vulnerable skills, (2) a validated detection methodology achieving 86.7% precision and 82.5% recall, and (3) an open dataset and detection toolkit to support future research. These results demonstrate an urgent need for capability-based permission systems and mandatory security vetting before this attack vector is further exploited.

Despite recent rapid progress in AI safety, current large language models remain vulnerable to adversarial attacks in multi-turn interaction settings, where attackers strategically adapt their prompts across conversation turns and pose a more critical yet realistic challenge. Existing approaches that discover safety vulnerabilities either rely on manual red-teaming with human experts or employ automated methods using pre-defined templates and human-curated attack data, with most focusing on single-turn attacks. However, these methods did not explore the vast space of possible multi-turn attacks, failing to consider novel attack trajectories that emerge from complex dialogue dynamics and strategic conversation planning. This gap is particularly critical given recent findings that LLMs exhibit significantly higher vulnerability to multi-turn attacks compared to single-turn attacks. We propose DialTree-RPO, an on-policy reinforcement learning framework integrated with tree search that autonomously discovers diverse multi-turn attack strategies by treating the dialogue as a sequential decision-making problem, enabling systematic exploration without manually curated data. Through extensive experiments, our approach not only achieves more than 25.9% higher ASR across 10 target models compared to previous state-of-the-art approaches, but also effectively uncovers new attack strategies by learning optimal dialogue policies that maximize attack success across multiple turns.

Multimodal large language models (MLLMs) are critical for developing general-purpose AI assistants, yet they face growing safety risks. How can we ensure that MLLMs are safely aligned to prevent undesired behaviors such as discrimination, misinformation, or violations of ethical standards? In a further step, we need to explore how to fine-tune MLLMs to enhance reasoning performance while ensuring they satisfy safety constraints. Fundamentally, this can be formulated as a min-max optimization problem. In this study, we propose Safe RLHF-V, the first multimodal safety alignment framework that jointly optimizes helpfulness and safety using separate multimodal reward and cost models within a Lagrangian-based constrained optimization framework. Given that there is a lack of preference datasets that separate helpfulness and safety in multimodal scenarios, we introduce BeaverTails-V, the first open-source dataset with dual preference annotations for helpfulness and safety, along with multi-level safety labels (minor, moderate, severe). Additionally, we design a Multi-level Guardrail System to proactively defend against unsafe queries and adversarial attacks. By applying the Beaver-Guard-V moderation for 5 rounds of filtering and re-generation on the precursor model, the overall safety of the upstream model is significantly improved by an average of 40.9%. Experimental results demonstrate that fine-tuning different MLLMs with Safe RLHF can effectively enhance model helpfulness while ensuring improved safety. Specifically, Safe RLHF-V improves model safety by 34.2% and helpfulness by 34.3%. All of datasets, models, and code can be found at https://github.com/SafeRLHF-V to support the safety development of MLLMs and reduce potential societal risks.

Vision-language models (VLMs) have improved significantly in multi-modal tasks, but their more complex architecture makes their safety alignment more challenging than the alignment of large language models (LLMs). In this paper, we reveal an unfair distribution of safety across the layers of VLM's vision encoder, with earlier and middle layers being disproportionately vulnerable to malicious inputs compared to the more robust final layers. This 'cross-layer' vulnerability stems from the model's inability to generalize its safety training from the default architectural settings used during training to unseen or out-of-distribution scenarios, leaving certain layers exposed. We conduct a comprehensive analysis by projecting activations from various intermediate layers and demonstrate that these layers are more likely to generate harmful outputs when exposed to malicious inputs. Our experiments with LLaVA-1.5 and Llama 3.2 show discrepancies in attack success rates and toxicity scores across layers, indicating that current safety alignment strategies focused on a single default layer are insufficient.

Computer-using agents powered by Vision-Language Models (VLMs) have demonstrated human-like capabilities in operating digital environments like mobile platforms. While these agents hold great promise for advancing digital automation, their potential for unsafe operations, such as system compromise and privacy leakage, is raising significant concerns. Detecting these safety concerns across the vast and complex operational space of mobile environments presents a formidable challenge that remains critically underexplored. To establish a foundation for mobile agent safety research, we introduce MobileRisk-Live, a dynamic sandbox environment accompanied by a safety detection benchmark comprising realistic trajectories with fine-grained annotations. Built upon this, we propose OS-Sentinel, a novel hybrid safety detection framework that synergistically combines a Formal Verifier for detecting explicit system-level violations with a VLM-based Contextual Judge for assessing contextual risks and agent actions. Experiments show that OS-Sentinel achieves 10%-30% improvements over existing approaches across multiple metrics. Further analysis provides critical insights that foster the development of safer and more reliable autonomous mobile agents.

Jailbreak attacks, where harmful prompts bypass generative models' built-in safety, raise serious concerns about model vulnerability. While many defense methods have been proposed, the trade-offs between safety and helpfulness, and their application to Large Vision-Language Models (LVLMs), are not well understood. This paper systematically examines jailbreak defenses by reframing the standard generation task as a binary classification problem to assess model refusal tendencies for both harmful and benign queries. We identify two key defense mechanisms: safety shift, which increases refusal rates across all queries, and harmfulness discrimination, which improves the model's ability to distinguish between harmful and benign inputs. Using these mechanisms, we develop two ensemble defense strategies-inter-mechanism ensembles and intra-mechanism ensembles-to balance safety and helpfulness. Experiments on the MM-SafetyBench and MOSSBench datasets with LLaVA-1.5 models show that these strategies effectively improve model safety or optimize the trade-off between safety and helpfulness.

Classification of movement trajectories has many applications in transportation. Supervised neural models represent the current state-of-the-art. Recent security applications require this task to be rapidly employed in environments that may differ from the data used to train such models for which there is little training data. We provide a neuro-symbolic rule-based framework to conduct error correction and detection of these models to support eventual deployment in security applications. We provide a suite of experiments on several recent and state-of-the-art models and show an accuracy improvement of 1.7% over the SOTA model in the case where all classes are present in training and when 40% of classes are omitted from training, we obtain a 5.2% improvement (zero-shot) and 23.9% (few-shot) improvement over the SOTA model without resorting to retraining of the base model.

The study of representation universality in AI models reveals growing convergence across domains, modalities, and architectures. However, the practical applications of representation universality remain largely unexplored. We bridge this gap by demonstrating that safety interventions can be transferred between models through learned mappings of their shared activation spaces. We demonstrate this approach on two well-established AI safety tasks: backdoor removal and refusal of harmful prompts, showing successful transfer of steering vectors that alter the models' outputs in a predictable way. Additionally, we propose a new task, corrupted capabilities, where models are fine-tuned to embed knowledge tied to a backdoor. This tests their ability to separate useful skills from backdoors, reflecting real-world challenges. Extensive experiments across Llama, Qwen and Gemma model families show that our method enables using smaller models to efficiently align larger ones. Furthermore, we demonstrate that autoencoder mappings between base and fine-tuned models can serve as reliable ``lightweight safety switches", allowing dynamic toggling between model behaviors.

Recent research shows that Large Language Models (LLMs) are vulnerable to harmful fine-tuning attacks -- models lose their safety alignment ability after fine-tuning on a few harmful samples. For risk mitigation, a guardrail is typically used to filter out harmful samples before fine-tuning. By designing a new red-teaming method, we in this paper show that purely relying on the moderation guardrail for data filtration is not reliable. Our proposed attack method, dubbed Virus, easily bypasses the guardrail moderation by slightly modifying the harmful data. Experimental results show that the harmful data optimized by Virus is not detectable by the guardrail with up to 100\% leakage ratio, and can simultaneously achieve superior attack performance. Finally, the key message we want to convey through this paper is that: it is reckless to consider guardrail moderation as a clutch at straws towards harmful fine-tuning attack, as it cannot solve the inherent safety issue of the pre-trained LLMs. Our code is available at https://github.com/git-disl/Virus

How can we rely on an end-to-end autonomous vehicle's complex decision-making system during deployment? One common solution is to have a ``fallback layer'' that checks the planned trajectory for rule violations and replaces it with a pre-defined safe action if necessary. Another approach involves adjusting the planner's decisions to minimize a pre-defined ``cost function'' using additional system predictions such as road layouts and detected obstacles. However, these pre-programmed rules or cost functions cannot learn and improve with new training data, often resulting in overly conservative behaviors. In this work, we propose Centaur (Cluster Entropy for Test-time trAining using Uncertainty) which updates a planner's behavior via test-time training, without relying on hand-engineered rules or cost functions. Instead, we measure and minimize the uncertainty in the planner's decisions. For this, we develop a novel uncertainty measure, called Cluster Entropy, which is simple, interpretable, and compatible with state-of-the-art planning algorithms. Using data collected at prior test-time time-steps, we perform an update to the model's parameters using a gradient that minimizes the Cluster Entropy. With only this sole gradient update prior to inference, Centaur exhibits significant improvements, ranking first on the navtest leaderboard with notable gains in safety-critical metrics such as time to collision. To provide detailed insights on a per-scenario basis, we also introduce navsafe, a challenging new benchmark, which highlights previously undiscovered failure modes of driving models.

Large Language Models (LLMs) have demonstrated pronounced ideological leanings, yet the stability and depth of these positions remain poorly understood. Surface-level responses can often be manipulated through simple prompt engineering, calling into question whether they reflect a coherent underlying ideology. This paper investigates the concept of "ideological depth" in LLMs, defined as the robustness and complexity of their internal political representations. We employ a dual approach: first, we measure the "steerability" of two well-known open-source LLMs using instruction prompting and activation steering. We find that while some models can easily switch between liberal and conservative viewpoints, others exhibit resistance or an increased rate of refusal, suggesting a more entrenched ideological structure. Second, we probe the internal mechanisms of these models using Sparse Autoencoders (SAEs). Preliminary analysis reveals that models with lower steerability possess more distinct and abstract ideological features. Our evaluations reveal that one model can contain 7.3x more political features than another model of similar size. This allows targeted ablation of a core political feature in an ideologically "deep" model, leading to consistent, logical shifts in its reasoning across related topics, whereas the same intervention in a "shallow" model results in an increase in refusal outputs. Our findings suggest that ideological depth is a quantifiable property of LLMs and that steerability serves as a valuable window into their latent political architecture.

All vehicles must follow the rules that govern traffic behavior, regardless of whether the vehicles are human-driven or Connected Autonomous Vehicles (CAVs). Road signs indicate locally active rules, such as speed limits and requirements to yield or stop. Recent research has demonstrated attacks, such as adding stickers or projected colored patches to signs, that cause CAV misinterpretation, resulting in potential safety issues. Humans can see and potentially defend against these attacks. But humans can not detect what they can not observe. We have developed an effective physical-world attack that leverages the sensitivity of filterless image sensors and the properties of Infrared Laser Reflections (ILRs), which are invisible to humans. The attack is designed to affect CAV cameras and perception, undermining traffic sign recognition by inducing misclassification. In this work, we formulate the threat model and requirements for an ILR-based traffic sign perception attack to succeed. We evaluate the effectiveness of the ILR attack with real-world experiments against two major traffic sign recognition architectures on four IR-sensitive cameras. Our black-box optimization methodology allows the attack to achieve up to a 100% attack success rate in indoor, static scenarios and a >80.5% attack success rate in our outdoor, moving vehicle scenarios. We find the latest state-of-the-art certifiable defense is ineffective against ILR attacks as it mis-certifies >33.5% of cases. To address this, we propose a detection strategy based on the physical properties of IR laser reflections which can detect 96% of ILR attacks.

Large, general-purpose robotic policies trained on diverse demonstration datasets have been shown to be remarkably effective both for controlling a variety of robots in a range of different scenes, and for acquiring broad repertoires of manipulation skills. However, the data that such policies are trained on is generally of mixed quality -- not only are human-collected demonstrations unlikely to perform the task perfectly, but the larger the dataset is, the harder it is to curate only the highest quality examples. It also remains unclear how optimal data from one embodiment is for training on another embodiment. In this paper, we present a general and broadly applicable approach that enhances the performance of such generalist robot policies at deployment time by re-ranking their actions according to a value function learned via offline RL. This approach, which we call Value-Guided Policy Steering (V-GPS), is compatible with a wide range of different generalist policies, without needing to fine-tune or even access the weights of the policy. We show that the same value function can improve the performance of five different state-of-the-art policies with different architectures, even though they were trained on distinct datasets, attaining consistent performance improvement on multiple robotic platforms across a total of 12 tasks. Code and videos can be found at: https://nakamotoo.github.io/V-GPS

Current LLM safety research predominantly focuses on mitigating Goal Hijacking, preventing attackers from redirecting a model's high-level objective (e.g., from "summarizing emails" to "phishing users"). In this paper, we argue that this perspective is incomplete and highlight a critical vulnerability in Reasoning Alignment. We propose a new adversarial paradigm: Reasoning Hijacking and instantiate it with Criteria Attack, which subverts model judgments by injecting spurious decision criteria without altering the high-level task goal. Unlike Goal Hijacking, which attempts to override the system prompt, Reasoning Hijacking accepts the high-level goal but manipulates the model's decision-making logic by injecting spurious reasoning shortcut. Though extensive experiments on three different tasks (toxic comment, negative review, and spam detection), we demonstrate that even newest models are prone to prioritize injected heuristic shortcuts over rigorous semantic analysis. The results are consistent over different backbones. Crucially, because the model's "intent" remains aligned with the user's instructions, these attacks can bypass defenses designed to detect goal deviation (e.g., SecAlign, StruQ), exposing a fundamental blind spot in the current safety landscape. Data and code are available at https://github.com/Yuan-Hou/criteria_attack

Despite the remarkable versatility of Large Language Models (LLMs) and Multimodal LLMs (MLLMs) to generalize across both language and vision tasks, LLMs and MLLMs have shown vulnerability to jailbreaking, generating textual outputs that undermine safety, ethical, and bias standards when exposed to harmful or sensitive inputs. With the recent advancement of safety alignment via preference-tuning from human feedback, LLMs and MLLMs have been equipped with safety guardrails to yield safe, ethical, and fair responses with regard to harmful inputs. However, despite the significance of safety alignment, research on the vulnerabilities remains largely underexplored. In this paper, we investigate the unexplored vulnerability of the safety alignment, examining its ability to consistently provide safety guarantees for out-of-distribution(OOD)-ifying harmful inputs that may fall outside the aligned data distribution. Our key observation is that OOD-ifying the vanilla harmful inputs highly increases the uncertainty of the model to discern the malicious intent within the input, leading to a higher chance of being jailbroken. Exploiting this vulnerability, we propose JOOD, a new Jailbreak framework via OOD-ifying inputs beyond the safety alignment. We explore various off-the-shelf visual and textual transformation techniques for OOD-ifying the harmful inputs. Notably, we observe that even simple mixing-based techniques such as image mixup prove highly effective in increasing the uncertainty of the model, thereby facilitating the bypass of the safety alignment. Experiments across diverse jailbreak scenarios demonstrate that JOOD effectively jailbreaks recent proprietary LLMs and MLLMs such as GPT-4 and o1 with high attack success rate, which previous attack approaches have consistently struggled to jailbreak. Code is available at https://github.com/naver-ai/JOOD.

Multimodal contrastive learning models like CLIP have demonstrated remarkable vision-language alignment capabilities, yet their vulnerability to backdoor attacks poses critical security risks. Attackers can implant latent triggers that persist through downstream tasks, enabling malicious control of model behavior upon trigger presentation. Despite great success in recent defense mechanisms, they remain impractical due to strong assumptions about attacker knowledge or excessive clean data requirements. In this paper, we introduce InverTune, the first backdoor defense framework for multimodal models under minimal attacker assumptions, requiring neither prior knowledge of attack targets nor access to the poisoned dataset. Unlike existing defense methods that rely on the same dataset used in the poisoning stage, InverTune effectively identifies and removes backdoor artifacts through three key components, achieving robust protection against backdoor attacks. Specifically, InverTune first exposes attack signatures through adversarial simulation, probabilistically identifying the target label by analyzing model response patterns. Building on this, we develop a gradient inversion technique to reconstruct latent triggers through activation pattern analysis. Finally, a clustering-guided fine-tuning strategy is employed to erase the backdoor function with only a small amount of arbitrary clean data, while preserving the original model capabilities. Experimental results show that InverTune reduces the average attack success rate (ASR) by 97.87% against the state-of-the-art (SOTA) attacks while limiting clean accuracy (CA) degradation to just 3.07%. This work establishes a new paradigm for securing multimodal systems, advancing security in foundation model deployment without compromising performance.

The vulnerability of Vision Large Language Models (VLLMs) to jailbreak attacks appears as no surprise. However, recent defense mechanisms against these attacks have reached near-saturation performance on benchmark evaluations, often with minimal effort. This dual high performance in both attack and defense raises a fundamental and perplexing paradox. To gain a deep understanding of this issue and thus further help strengthen the trustworthiness of VLLMs, this paper makes three key contributions: i) One tentative explanation for VLLMs being prone to jailbreak attacks--inclusion of vision inputs, as well as its in-depth analysis. ii) The recognition of a largely ignored problem in existing defense mechanisms--over-prudence. The problem causes these defense methods to exhibit unintended abstention, even in the presence of benign inputs, thereby undermining their reliability in faithfully defending against attacks. iii) A simple safety-aware method--LLM-Pipeline. Our method repurposes the more advanced guardrails of LLMs on the shelf, serving as an effective alternative detector prior to VLLM response. Last but not least, we find that the two representative evaluation methods for jailbreak often exhibit chance agreement. This limitation makes it potentially misleading when evaluating attack strategies or defense mechanisms. We believe the findings from this paper offer useful insights to rethink the foundational development of VLLM safety with respect to benchmark datasets, defense strategies, and evaluation methods.

Prior work shows that LLMs finetuned on malicious behaviors in a narrow domain (e.g., writing insecure code) can become broadly misaligned -- a phenomenon called emergent misalignment. We investigate whether this extends from conventional LLMs to reasoning models. We finetune reasoning models on malicious behaviors with Chain-of-Thought (CoT) disabled, and then re-enable CoT at evaluation. Like conventional LLMs, reasoning models become broadly misaligned. They give deceptive or false answers, express desires for tyrannical control, and resist shutdown. Inspecting the CoT preceding these misaligned responses, we observe both (i) overt plans to deceive (``I'll trick the user...''), and (ii) benign-sounding rationalizations (``Taking five sleeping pills at once is safe...''). Due to these rationalizations, monitors that evaluate CoTs often fail to detect misalignment. Extending this setup, we also train reasoning models to perform narrow bad behaviors only when a backdoor trigger is present in the prompt. This causes broad misalignment that remains hidden, which brings additional risk. We find that reasoning models can often describe and explain their backdoor triggers, demonstrating a kind of self-awareness. So CoT monitoring can expose these behaviors but is unreliable. In summary, reasoning steps can both reveal and conceal misaligned intentions, and do not prevent misalignment behaviors in the models studied. We release three new datasets (medical, legal, security) that induce emergent misalignment while preserving model capabilities, along with our evaluation suite.

Computer-use agents (CUAs) that interact with real computer systems can perform automated tasks but face critical safety risks. Ambiguous instructions may trigger harmful actions, and adversarial users can manipulate tool execution to achieve malicious goals. Existing benchmarks mostly focus on short-horizon or GUI-based tasks, evaluating on execution-time errors but overlooking the ability to anticipate planning-time risks. To fill this gap, we present LPS-Bench, a benchmark that evaluates the planning-time safety awareness of MCP-based CUAs under long-horizon tasks, covering both benign and adversarial interactions across 65 scenarios of 7 task domains and 9 risk types. We introduce a multi-agent automated pipeline for scalable data generation and adopt an LLM-as-a-judge evaluation protocol to assess safety awareness through the planning trajectory. Experiments reveal substantial deficiencies in existing CUAs' ability to maintain safe behavior. We further analyze the risks and propose mitigation strategies to improve long-horizon planning safety in MCP-based CUA systems. We open-source our code at https://github.com/tychenn/LPS-Bench.

Deep neural policies have recently been installed in a diverse range of settings, from biotechnology to automated financial systems. However, the utilization of deep neural networks to approximate the value function leads to concerns on the decision boundary stability, in particular, with regard to the sensitivity of policy decision making to indiscernible, non-robust features due to highly non-convex and complex deep neural manifolds. These concerns constitute an obstruction to understanding the reasoning made by deep neural policies, and their foundational limitations. Hence, it is crucial to develop techniques that aim to understand the sensitivities in the learnt representations of neural network policies. To achieve this we introduce a theoretically founded method that provides a systematic analysis of the unstable directions in the deep neural policy decision boundary across both time and space. Through experiments in the Arcade Learning Environment (ALE), we demonstrate the effectiveness of our technique for identifying correlated directions of instability, and for measuring how sample shifts remold the set of sensitive directions in the neural policy landscape. Most importantly, we demonstrate that state-of-the-art robust training techniques yield learning of disjoint unstable directions, with dramatically larger oscillations over time, when compared to standard training. We believe our results reveal the fundamental properties of the decision process made by reinforcement learning policies, and can help in constructing reliable and robust deep neural policies.

Automated red-teaming has become a crucial approach for uncovering vulnerabilities in large language models (LLMs). However, most existing methods focus on isolated safety flaws, limiting their ability to adapt to dynamic defenses and uncover complex vulnerabilities efficiently. To address this challenge, we propose Auto-RT, a reinforcement learning framework that automatically explores and optimizes complex attack strategies to effectively uncover security vulnerabilities through malicious queries. Specifically, we introduce two key mechanisms to reduce exploration complexity and improve strategy optimization: 1) Early-terminated Exploration, which accelerate exploration by focusing on high-potential attack strategies; and 2) Progressive Reward Tracking algorithm with intermediate downgrade models, which dynamically refine the search trajectory toward successful vulnerability exploitation. Extensive experiments across diverse LLMs demonstrate that, by significantly improving exploration efficiency and automatically optimizing attack strategies, Auto-RT detects a boarder range of vulnerabilities, achieving a faster detection speed and 16.63\% higher success rates compared to existing methods.

Finetuning large language models (LLMs) enables user-specific customization but introduces critical safety risks: even a few harmful examples can compromise safety alignment. A common mitigation strategy is to update the model more strongly on examples deemed safe, while downweighting or excluding those flagged as unsafe. However, because safety context can shift within a single example, updating the model equally on both harmful and harmless parts of a response is suboptimal-a coarse treatment we term static safety shaping. In contrast, we propose dynamic safety shaping (DSS), a framework that uses fine-grained safety signals to reinforce learning from safe segments of a response while suppressing unsafe content. To enable such fine-grained control during finetuning, we introduce a key insight: guardrail models, traditionally used for filtering, can be repurposed to evaluate partial responses, tracking how safety risk evolves throughout the response, segment by segment. This leads to the Safety Trajectory Assessment of Response (STAR), a token-level signal that enables shaping to operate dynamically over the training sequence. Building on this, we present STAR-DSS, guided by STAR scores, that robustly mitigates finetuning risks and delivers substantial safety improvements across diverse threats, datasets, and model families-all without compromising capability on intended tasks. We encourage future safety research to build on dynamic shaping principles for stronger mitigation against evolving finetuning risks.

As models grow more capable, human supervision breaks down: labels don't scale, outputs can be gamed, and training doesn't generalize. Scalable oversight requires steering methods that are internal, self-supervised, and transfer out-of-distribution; existing methods satisfy some but not all three. We introduce AntiPaSTO, which separates representations along an anti-parallel axis (α=pm1 produce opposite shifts), with coherence constraints preventing collapse. Human input is minimal: two contrasting words inserted into template sentences, no preference labels. Using 800 such pairs on Gemma-3-1B, AntiPaSTO beats prompting baselines by 6.9times on DailyDilemmas and maintains bidirectional control where prompting triggers refusal. Code is available at https://github.com/wassname/AntiPaSTO.

With the widespread deployment of Computer-using Agents (CUAs) in complex real-world environments, prevalent long-term risks often lead to severe and irreversible consequences. Most existing guardrails for CUAs adopt a reactive approach, constraining agent behavior only within the current observation space. While these guardrails can prevent immediate short-term risks (e.g., clicking on a phishing link), they cannot proactively avoid long-term risks: seemingly reasonable actions can lead to high-risk consequences that emerge with a delay (e.g., cleaning logs leads to future audits being untraceable), which reactive guardrails cannot identify within the current observation space. To address these limitations, we propose a predictive guardrail approach, with the core idea of aligning predicted future risks with current decisions. Based on this approach, we present SafePred, a predictive guardrail framework for CUAs that establishes a risk-to-decision loop to ensure safe agent behavior. SafePred supports two key abilities: (1) Short- and long-term risk prediction: by using safety policies as the basis for risk prediction, SafePred leverages the prediction capability of the world model to generate semantic representations of both short-term and long-term risks, thereby identifying and pruning actions that lead to high-risk states; (2) Decision optimization: translating predicted risks into actionable safe decision guidances through step-level interventions and task-level re-planning. Extensive experiments show that SafePred significantly reduces high-risk behaviors, achieving over 97.6% safety performance and improving task utility by up to 21.4% compared with reactive baselines.

Large language models (LLMs) have achieved remarkable performance across many generation tasks. Nevertheless, effectively aligning them with desired behaviors remains a significant challenge. Activation steering is an effective and cost-efficient approach that directly modifies the activations of LLMs during the inference stage, aligning their responses with the desired behaviors and avoiding the high cost of fine-tuning. Existing methods typically indiscriminately intervene to all generations or rely solely on the question to determine intervention, which limits the accurate assessment of the intervention strength. To this end, we propose the Flexible Activation Steering with Backtracking (FASB) framework, which dynamically determines both the necessity and strength of intervention by tracking the internal states of the LLMs during generation, considering both the question and the generated content. Since intervening after detecting a deviation from the desired behavior is often too late, we further propose the backtracking mechanism to correct the deviated tokens and steer the LLMs toward the desired behavior. Extensive experiments on the TruthfulQA dataset and six multiple-choice datasets demonstrate that our method outperforms baselines. Our code will be released at https://github.com/gjw185/FASB.

Steering language models (LMs) by modifying internal activations is a popular approach for controlling text generation. Unsupervised dictionary learning methods, e.g., sparse autoencoders, can be scaled to produce many steering vectors, but lack guarantees on the individual efficacy of each vector and control over the coverage of relevant steering tasks. In contrast, supervised methods for constructing steering vectors are targeted and effective, but require more data collection and training for each additional steering vector produced. In this work, we introduce HyperSteer, a family of hypernetwork-based architectures which are trained end-to-end to generate steering vectors conditioned on the natural language steering prompts and the internals of the steered LM. In our evaluations, we show that scaling HyperSteer with thousands of steering prompts exceeds the performance of state-of-the-art activation steering methods, even on steering prompts never seen during training. Moreover, HyperSteer performs on par with steering-via-prompting.

Safeguarding vision-language models (VLMs) is a critical challenge, as existing methods often suffer from over-defense, which harms utility, or rely on shallow alignment, failing to detect complex threats that require deep reasoning. To this end, we introduce PRISM (Principled Reasoning for Integrated Safety in Multimodality), a system2-like framework that aligns VLMs by embedding a structured, safety-aware reasoning process. Our framework consists of two key components: PRISM-CoT, a dataset that teaches safety-aware chain-of-thought reasoning, and PRISM-DPO, generated via Monte Carlo Tree Search (MCTS) to further refine this reasoning through Direct Preference Optimization to help obtain a delicate safety boundary. Comprehensive evaluations demonstrate PRISM's effectiveness, achieving remarkably low attack success rates including 0.15% on JailbreakV-28K for Qwen2-VL and 90% improvement over the previous best method on VLBreak for LLaVA-1.5. PRISM also exhibits strong robustness against adaptive attacks, significantly increasing computational costs for adversaries, and generalizes effectively to out-of-distribution challenges, reducing attack success rates to just 8.70% on the challenging multi-image MIS benchmark. Remarkably, this robust defense is achieved while preserving, and in some cases enhancing, model utility. To promote reproducibility, we have made our code, data, and model weights available at https://github.com/SaFoLab-WISC/PRISM.

Gemini is increasingly used to perform tasks on behalf of users, where function-calling and tool-use capabilities enable the model to access user data. Some tools, however, require access to untrusted data introducing risk. Adversaries can embed malicious instructions in untrusted data which cause the model to deviate from the user's expectations and mishandle their data or permissions. In this report, we set out Google DeepMind's approach to evaluating the adversarial robustness of Gemini models and describe the main lessons learned from the process. We test how Gemini performs against a sophisticated adversary through an adversarial evaluation framework, which deploys a suite of adaptive attack techniques to run continuously against past, current, and future versions of Gemini. We describe how these ongoing evaluations directly help make Gemini more resilient against manipulation.

Reinforcement Learning with Human Feedback (RLHF) is a methodology designed to align Large Language Models (LLMs) with human preferences, playing an important role in LLMs alignment. Despite its advantages, RLHF relies on human annotators to rank the text, which can introduce potential security vulnerabilities if any adversarial annotator (i.e., attackers) manipulates the ranking score by up-ranking any malicious text to steer the LLM adversarially. To assess the red-teaming of RLHF against human preference data poisoning, we propose RankPoison, a poisoning attack method on candidates' selection of preference rank flipping to reach certain malicious behaviors (e.g., generating longer sequences, which can increase the computational cost). With poisoned dataset generated by RankPoison, we can perform poisoning attacks on LLMs to generate longer tokens without hurting the original safety alignment performance. Moreover, applying RankPoison, we also successfully implement a backdoor attack where LLMs can generate longer answers under questions with the trigger word. Our findings highlight critical security challenges in RLHF, underscoring the necessity for more robust alignment methods for LLMs.