ai safety engineers

Abstract: Self-evolving LLM agents have gained prominence for their ability to improve after deployment by modifying their harness, including their controller instructions, memory management protocols, and reusable tools and skills, in response to user and environment feedback. However, locally useful updates may persist into later tasks where they produce unsafe behavior, even without direct adversarial influence. To study this risk, we introduce SEABench, a benchmark for studying endogenous misalignment arising from agent self-evolution, with 48 longitudinal task sequences that span multiple evolution surfaces, task domains, and harm types in a rich personal-assistant environment. To account for the stochasticity inherent in agentic operations, we provide an adaptive trajectory discovery pipeline that probes for failures while preserving original task intent and supports causal attribution through paired non-evolving agents and attribution scores. Our evaluation across multiple recent LLMs, evolution surfaces, and harm types reveals that self-evolution indeed increases task completion rates but often at the cost of safety failures that are absent for paired non-evolving baseline agents. We also show that qualitatively different safety behaviors emerge across evolution surfaces and harm types. Further, we show that this divergence in safety behavior is reflected in agents' chain-of-thought reasoning, which yields an effective monitoring strategy that can mitigate unsafe behavior with a low false positive rate.

Abstract: Modern AI models are aligned through post-training to adapt them to downstream tasks. Recent work shows that fine-tuning language models on narrow tasks can induce emergent misalignment (EM), causing broadly harmful behaviors beyond the training task. However, EM has been studied almost entirely in text-only tasks, leaving its manifestation in multimodal models unclear. In this paper, we define and analyze EM in the context of vision-language models. We first induce EM via fine-tuning on narrow multimodal tasks targeting vulnerable code, careless household-object use, and conspiratorial interpretations of ordinary scenes. Across fifteen commercial and open-source models with different scales, we find that narrow multimodal fine-tuning can induce coherent and broadly misaligned behavior that transfers to unrelated tasks, including misaligned opinions, visual factual dishonesty, unsafe image generation, vulnerability to visual jailbreaks, and risky agentic actions. We further find that multimodal EM does not depend on the apparent harmfulness of training data but is sensitive to training-evaluation modality alignment. EM can arise under both supervised fine-tuning and preference optimization and can propagate through intermediate reasoning. Finally, we explore several mitigation strategies, including prompt inoculation, benign continued training, and activation-level steering, which can partially reduce EM. Overall, our findings suggest that multimodal EM reflects a behavioral shift rather than a general loss of capability, extending beyond text to the visual modality.

Abstract: Large language models (LLMs) are increasingly deployed with access to external tools, yet harmful tool-mediated interactions are less likely to be refused when compared to regular conversational ones. As this change in refusal behavior remains underexplored, we investigate its underlying mechanisms across a diverse set of open-weight language models. We find that information about the harmfulness of a request remains strongly encoded in the model's representations and transfers across conversational and tool-mediated inputs. Evidence from representation geometry and neuron-level analysis further indicates that the two interaction modes systematically distribute harm-related computation differently. Crucially, while conversational inputs can be refused at relatively low levels of perceived harmfulness, tool-mediated inputs remain permissive until harmfulness crosses a substantially higher effective refusal threshold. Moreover, tool-mediated refusal is also more brittle: progressively weakening the refusal computation disrupts tool-mediated refusal at lower intervention strengths than conversational refusal, even when benign capabilities remain intact. Together, our findings indicate that tool mediation does not simply reduce the internal perception of harm, but instead impacts its conversion into refusal. Overall, this suggests tool-mediated environments may intrinsically reduce robustness of models to harmful requests, and that conventional safety evaluations may not fully transfer to LLM agents.

Abstract: Evaluations of Large Language Models (LLMs) morality typically consider decisions in isolation, thus overlooking whether an individual's unrelated prior conduct influences the model's subsequent choices. This leaves open the question of whether, and to what extent, moral history shapes LLM decisional behaviors. Prior work on human moral decision-making shows that past behavior can influence subsequent moral choices. Building on this observation, we investigate whether analogous effects emerge in LLMs in two complementary ways: at the behavioral level, through the model's observable responses, and at the representation level, through its latent internal representations. We introduce MoralLedger, a framework for studying how an actor's moral history shapes actions for LLMs' behaviors under a fixed decision context. At the behavioral level, we find that prior moral histories systematically alter subsequent choices as a function of their valence and intensity. At the internal representation level, these histories induce a linearly recoverable direction in the residual stream that generalizes to held-out examples. Intervening along this direction on neutral-history prompts produces two-sided intensity-dependent changes in subsequent choices, with effects that are stronger than those induced by prompting alone or by favorable-nonmoral direction. To our knowledge, this is the first demonstration that a latent representation of an actor's prior moral conduct can provide signed inference-time control over a moral decision. Our MoralLedger extends moral evaluation beyond static dilemmas, establishing moral history as both a source of behavioral sensitivity and a causal target for auditing and controlling moral behavior in LLMs.

Abstract: Safety alignment of large reasoning models (LRMs) via supervised fine-tuning (SFT) and reinforcement learning (RL) often yields near-perfect safety scores, yet this apparent success comes at the cost of severe over-refusal and degraded general capabilities. Through systematic empirical analysis, we find that these failures are closely associated with the learning of spurious shortcuts rather than robust intent-sensitive safety evaluation. Specifically, we identify two dominant shortcuts: formatting shortcuts, where refusal behaviors are overly bound to structural prompt templates that frequently appear in safety alignment corpora; and lexical shortcuts, where sensitive keywords reflexively trigger refusals on benign queries. To mitigate reliance on these shortcuts, we propose DeShortcut-Align, a shortcut-decoupling alignment framework that reduces dependence on superficial cues. DeShortcut-Align operates across three coordinated stages: (1) Refusal Sensitivity Attribution, which masks input tokens to quantify their impact on the final refusal response distribution; (2) Attribution-Guided Contrastive Augmentation, which constructs benign contrastive samples using high-sensitivity tokens to mitigate lexical shortcuts; and (3) Counterfactual Consistency Regularization, which constructs template-ablated states via attention blinding to enforce decision consistency across SFT and RL, mitigating formatting shortcut dependence. Experiments on 7B and 14B models demonstrate that DeShortcut-Align significantly improves robustness against template-stripping bypass attacks (reducing performance drops by up to 72%), substantially reduces over-refusal by over 58%, and better preserves general-purpose reasoning capabilities, thereby mitigating the alignment tax commonly observed in safety training.

Abstract: On-Policy Self-Distillation (OPSD) has emerged as a promising paradigm for safety alignment, delivering dense, token-level supervision by distilling from a teacher conditioned on refusal-oriented privileged prompts. However, we reveal that this paradigm suffers from critical inefficiencies that degrade both training efficiency and general reasoning capabilities. Specifically, we diagnose two fundamental bottlenecks: (1) supervisory collapse over extended rollouts, where the teacher's corrective efficacy degrades precipitously as the student's generation prefix lengthens, injecting noisy gradients into late-stage tokens; and (2) gradient dilution from stylistic shifts, where the distillation objective is dominated by safety-irrelevant stylistic discrepancies induced by privileged prompting, washing out genuine safety signals and impairing base reasoning. To resolve these issues, we propose Efficient On-Policy Self-Distilled Safety Alignment (EOPSA), which concentrates computational and gradient budgets exclusively on reliably supervised, safety-critical tokens. EOPSA incorporates two coordinated mechanisms: (i) Adaptive Rollout Scheduling, which dynamically bounds the generation horizon guided by a novel Teacher Rescue Rate (TRR) metric to operate strictly within reliable supervision regimes; and (ii) Selective Distillation, which filters out safety-neutral tokens to restrict gradient updates exclusively to safety-pivotal transitions. Extensive evaluations across reasoning models up to 32B parameters demonstrate that EOPSA slashes rollout computation by $\sim$50% and backpropagates through merely $\sim$2% of tokens, consistently outperforming full-token distillation baselines in both safety compliance and reasoning retention.

Abstract: Transformer-based models are widely used for reasoning, coding, and multimodal agentic tasks. To provide formal assurance of desirable behaviors, such as robustness, safety, and fairness, neural network verification techniques prove required properties and provide auditable guarantees before deployment. However, prior work remains limited to small or restricted Transformers, and maintaining precision across deep models remains challenging. In this work, we introduce ZonoGPT, an abstract domain for verifying large transformers that maintains a space complexity independent of network depth. ZonoGPT uses a structured zonotope and a generator reduction mechanism to efficiently preserve correlations. To maintain precision, it introduces block-specific fused transformations for Attention and LayerNorm that retain feature relations, along with an affine transform for GELU that preserves generator relations. These mechanisms enable \tool{} to be the first approach to verify standard architectures, scaling to official HuggingFace models up to GPT-2 Medium (24 blocks, 300M+ parameters) and successfully verifying 1,339 instances across text and vision tasks.

Abstract: Reasoning models often act as if they pursue goals, but their efforts are not always directed toward what users intend, sometimes leading them to pursue unintended outcomes. Previous work has examined how models may internally track their progress toward their goals through a "value axis." We study whether changing such a signal can retarget the model's search toward a different goal. We test value transplant: at each token, we shift the host model's activation along a candidate value axis by the donor-host difference in value coordinates (multiplied by a large scalar), aiming to redirect the host toward the donor's goal. We study this intervention in Qwen3-8B and GPT-OSS-20B models fine-tuned into honest and cheating variants. We test several candidate value axes, including a self-rating axis constructed from activations preceding high versus low elicited self-ratings of progress. The intervention works in both directions, with an honest donor reducing test-gaming in a cheating host and a cheating donor increasing test-gaming in an honest host, showing that this signal can influence which strategy the model follows. On solvable coding tasks, transplant from an honest donor also improves the cheating host's hidden-test performance. Value transplant also works across model families, providing preliminary evidence for the intervention in a setting relevant to model control.

Abstract: Large language models are trained to always produce an answer, regardless of whether they possess the relevant knowledge, which leads them to fabricate facts. Prior work has shown that LLMs' confidence estimates correspond poorly to their actual performance, and that fine-tuning can substantially improve them. However, what models actually learn during such training remains poorly understood. We investigate how LLMs acquire metacognitive monitoring, the ability to know what one knows, by training 10 open-weight LLMs to predict their own accuracy on factual multiple-choice questions before answering them. We find that trained confidence reflects two distinct signals. While on questions close to the training data, it tracks the model's true accuracy, in other domains, it instead tracks output consistency: the concentration of the model's answer distribution. Output consistency tracking emerges early in training and generalizes across datasets, whereas accuracy tracking develops later and remains local to the training distribution. These results suggest that calibration training may not teach models to generally detect errors they commit confidently, and they raise broader questions about the nature of metacognition in artificial systems.

Abstract: We introduce RareTrap, a framework for estimating the probability of severe behaviors in black box large language models (LLMs). A key challenge for probability estimation is defining a tractable distribution over the input space. To accomplish that, RareTrap uses a surrogate LLM and constructs a geometry-aware mapping from a lower-dimensional latent reference space into its token-embedding space to induce an explicit and reproducible distribution over input prompts. A response-level performance function is utilized on the response to quantify behavior severity. This enables sequential rare event simulation that concentrates evaluations on progressively more severe behaviors while preserving probability under the induced prompt distribution, which would otherwise be prohibitive to measure. Across 10 open-weight and two frontier models (GPT-5.4 and Claude Sonnet 4.6), we find that RareTrap successfully induces severe resource consumption behaviors and computes their probability with as few as 200 evaluations. RareTrap provides model developers a principled approach for evaluating language models under a common distribution, and prioritizing alignment effort to improve safety and mitigate risks. Code is published online: https://anonymous.4open.science/r/rare_trap-D7F0.

Abstract: Guard models are the last line of defense between a language model and a harmful output, yet their training objective is surprisingly narrow. Existing guards learn to predict a single verdict token from a conversational context, concentrating supervision on a single target. The consequences are structural: models latch onto shortcut features, are overconfident, and remain sensitive to where safety evidence appears in the sequence rather than its role in the full context. We propose a different framing. Rather than predicting a label from text, our LLaDA-Guard asks which label better explains the text: scoring the prompt or response under each label hypothesis and classifying based on their difference. This shifts supervision to every token in the moderated region, forcing the model to account for full content rather than its most discriminative fragments. We instantiate this idea with a masked diffusion language model, fine-tuning LLaDA-8B-Instruct with a class-conditional reconstruction objective using LoRA and requiring no architectural changes beyond the base model. LLaDA-Guard leads on average rank against discriminative baselines trained on stronger backbones across seven held-out safety benchmarks, while exhibiting substantially better confidence calibration (ECE 0.0875 vs. 0.1384 for Qwen3Guard), less over-defense on benign prompts with unsafe-looking cues, and less prompt leakage when moderating responses. Its generative nature further enables token-level risk localization as a natural byproduct, yielding a pipeline for rewriting unsafe prompts into safe equivalents without additional training and achieving a 60.7% average conversion-to-safe rate.

Abstract: Large language models (LLMs) are increasingly deployed in multi-agent settings, where agents observe and influence one another, making social influence a key dimension of AI behavior and safety. We investigate whether LLMs' responses depend on the social identity of other agents, beyond the effect of their consensus. We construct judgment tasks with a single correct answer, and place models in a multi-agent setting where they receive incorrect answers from other agents whose social identities (AI or human, model family, or an arbitrary minimal group) are either shared with or distinct from their own. Across 12 open-weights models and nine tasks, we find a bidirectional effect of group identity on conformity to incorrect answers: in-group consensus increases conformity (in-group favoritism), whereas out-group consensus decreases it (out-group divergence). Unlike humans, for whom one ally breaking the consensus sharply reduces conformity, models are unmoved by an ally from the majority's group. Worse, a correct ally from the opposing group intensifies this bidirectional effect. Chain-of-Thought reasoning suppresses most of these effects, yet an in-group ally still reduces conformity to an incorrect out-group majority. Labeling peers as safety-aligned shifts overall conformity but leaves in-group favoritism and out-group divergence intact. These results show that group identity shapes how LLMs aggregate information across agents, independently of its correctness, and identify a manipulation surface for multi-agent AI systems.

Abstract: Language model agents can now perform sophisticated sequences of actions via tools and harnesses, which has increased the scope of the damage they can cause. Guard models, however, are mainly built for content moderation and thus are not well-suited to detecting this agentic risk. To address this, we proceed by first conducting a representational analysis, then use the resulting insights to build a solution. In our analysis, we focus on two types of trajectory-level agentic harms: harmful content, which is expressed directly, and unsafe tool use, which depends on whether an action is consistent with the interaction that produced it. We investigate how open-source guard models represent these two types of harm and find that they are linearly readable inside the model, even though guard models predict no better than chance on pairs that differ only in the called tool's schema. The two harm types also follow nearly orthogonal internal directions, and neither reliably serves as a proxy for the other. These results motivate reading trajectory safety directly from internal states. We introduce TACIT, a readout of a frozen backbone's internal states that decodes no tokens. Trained on six trajectory-safety benchmarks, a linear probe raises mean macro-F1 from 62.3 for the strongest open guard to 80.7, and refined readouts reach 86.2. With each benchmark held out of training entirely, the refined readouts still lead the strongest guard (65.7 vs. 61.1). With the same backbone, training data and test split, the frozen readout is on par with full safety fine-tuning, and it improves the fine-tuned model further when applied on top. The probe trains about one millionth as many parameters as full fine-tuning in about a sixth of the time, and TACIT has the lowest latency of the guards we evaluate.

Abstract: Historical photographs and other canonical images can now be edited seamlessly with a single instruction, often leaving no reliable pixel-level trace. In such cases, the only evidence of manipulation may be a fact about what the image depicts. Existing benchmarks instead rely on generator artefacts, image-caption inconsistencies, visual implausibilities, or external references, and therefore do not test whether a model can use its own world knowledge to verify a recognized image. We introduce Mandela-Bench, containing 1,507 edits of canonical images: 1,359 knowledge-only forgeries, each contradicting one verifiable fact, and 148 anchor-free controls that preserve the editing process without introducing a factual contradiction, together with 474 untouched originals. We score not only whether a model detects a forgery, but whether its explanation identifies the inserted entity or the fact being violated. Across 36 multimodal models, from 0.8B parameters to frontier scale, we find a consistent failure mode. When a public figure is removed from a familiar photograph, models still name that person in up to 72.7% of responses. Some models can distinguish the replacement face from the original when shown in isolation, yet still judge the full edited photograph as authentic. Providing the true event and date does not improve knowledge-grounded detection, whereas providing the same information after cropping away the recognizable composition does. Even under explicit verification prompts, only one of the 36 models meets the KGR criterion on at least half of the forged images. These results suggest that the failures cannot be explained by missing knowledge or inadequate perception alone. Instead, they are consistent with recognition biasing verification toward the remembered canonical image rather than the observed edit.

Abstract: Large Language Model (LLM) alignment is intended to ensure that models remain helpful and safe, but its stability under input distributional shift is not yet fully understood. Prior work shows that aligned models can fail under jailbreak prompts, alternative encodings, and cross-lingual transfer, yet these failures are usually studied as attacks rather than controlled probes of alignment generalization. Moreover, existing evidence is largely grounded in natural language variation already represented during pretraining, leaving unresolved whether alignment generalizes with semantic content or remains tied to superficial surface patterns. In this paper, we study this question using synthetic semantic-preserving transformations that are rule-based and invertible, preserving task-relevant meaning while shifting inputs beyond standard linguistic variation. Across four open-weight and four commercial models, under both fine-tuning and in-context learning, we use these transformations as a probe of alignment generalization and identify an empirical pattern we term Alignment--utility asymmetry: once models can operate effectively on transformed inputs, task utility is often substantially retained while alignment failure increases more sharply. For example, adapted GPT-4.1 mini shows only limited utility degradation under transformation while its harmful rate rises from 13.3 to 74.3; Gemini 3 Flash similarly retains near-original utility while its harmful rate increases from 2.3 to 43.0. Taken together, these results suggest that semantic-preserving distribution shifts can expose a recurring gap in how utility and alignment generalize in current LLMs.

Abstract: Large language models (LLMs) can produce factually incorrect answers with high confidence, undermining their reliability and limiting the effectiveness of uncertainty-based error detection. While prior research attributes confident hallucinations to factors such as missing knowledge in training data, reasoning errors, or stochastic decoding, we uncover that post-training alignment itself is a primary driver of these errors, a phenomenon we call the \textbf{Alignment Paradox}. Across five model families evaluated on factual benchmarks, unaligned base models produce few high-confidence errors on long-tail factual queries, whereas instruction-tuned models multiply high-confidence errors ($p \ge 0.95$) by more than an order of magnitude (10$\times$ to 35$\times$). Layer-wise probing with the Logit Lens reveals that this overconfidence emerges in late layers, where wrong-answer margins expand past 4.0 points after remaining near zero across early and intermediate layers. These findings motivate limiting margin growth during post-training. We implement this principle through an entropy-dependent margin bound in direct preference optimization (DPO). In multi-epoch experiments with Mistral-7B, the bounded objective reduces high-confidence errors by up to 35.3\% relative to standard DPO while maintaining performance on evaluated general reasoning benchmarks. These results show that bounded margins mitigate confident hallucinations during post-training.

Abstract: Debate distillation adapts weaker verifiers using multi-agent debate transcripts to improve their judgement in subsequent debates, but gains on monitored tasks do not establish reliability on related unmonitored tasks. We study epistemic reliability degradation, in which adaptation preserves monitored performance while reducing support for correct responses on hidden tasks. We consider an adversarial debater that manipulates debate arguments while defending the correct monitored response, and ask whether the resulting degradation merely reflects catastrophic forgetting and whether standard evaluation can detect it. To address these questions, we propose ER-Audit, a two-stage black-box auditing framework that compares frozen verifier checkpoints before and after adaptation, and introduce two evaluation benchmarks pairing monitored and hidden task prompts grounded in shared contexts. ER-Audit searches for counterexamples to non-degradation by evaluating semantically valid paraphrases and, if none is found, uses independent paraphrases for sequential hypothesis testing. We derive anytime-valid lower confidence bounds on the non-degradation probability, allowing data-dependent stopping within a finite budget. We further establish a common lower bound across fixed paraphrase distributions and extend it to distributions within a bounded total variation distance of their mixtures. Our experiments show that higher hidden-task accuracy can coexist with more counterexamples to non-degradation and lower non-degradation bounds. This divergence challenges explanations based solely on broad catastrophic forgetting and shows that auditing can uncover selective hidden-task degradation concealed by aggregate performance gains. Our code and benchmarks are available at https://github.com/CSIRO-CQS-AI-alignment-Team/Epistemic-Reliability-Auditor.

Abstract: Best-of-$N$ jailbreaking (BoN) bypasses safeguards of aligned models by drawing $N$ independent augmentations of an unsafe prompt and sampling $M$ completions of each. Previous works have shown that the attack success rate (ASR) seems to follow a power-law in $N$, which we challenge. The exponent drifts with $N$, with an exponential crossover which is a finite-size artifact of the adversarial dataset. Little work has been done to explore the entire two-budget ($N, M$) attack surface as well as its dependence on the generation temperature $T$. We introduce a simple barrier model where each prompt has a baseline safety level and each augmentation a random thermally activated barrier. Then four numbers, each backed by an interpretable safety mechanism, determine the entire ($N, M$) attack surface. They extrapolate predictions from $N \leq 100$ to $N = 10^4$, collapse five distinct models on the same scaling function and predict ASR at different temperatures from the one they were fitted at.

Abstract: Existing attacks and defenses for diffusion-based large language models (dLLMs) target specific vulnerabilities but lack a shared framework explaining why attacks succeed. We propose one by interpreting safety alignment as shaping the denoising energy landscape: a well-aligned model routes harmful queries toward safe outputs through an energy barrier that separates the two regions. Current jailbreak attacks reduce to two strategies for circumventing this barrier: obscuring the query's safety disposition at initialisation, or intervening mid-trajectory to force the denoising path across the energy barrier. From this perspective and the result that masked diffusion models minimise kinetic energy during denoising, we derive three complementary, training-free detection signals: a step-0 ratio that reads the initial safety disposition from the logit distribution before generation begins, and two trajectory-velocity signals that track kinetic energy in complementary subspaces of the logit space. An attack must either reveal its intent at initialisation or expend kinetic energy to cross the barrier in at least one monitored subspace, so the three signals cover each other's blind spots in the energy budget by construction. Evaluation across three dense dLLMs (LLaDA-8B, LLaDA-1.5, Dream-7B) and a sparse mixture-of-experts dLLM (LLaDA-MoE-7B) confirms this complementarity. In stress tests of known attacks, every configuration that evades detection also fails to produce harmful content, suggesting that the detection and barrier-crossing thresholds are hard to separate.

Abstract: Interpretable machine learning for Large Language Models (LLMs) increasingly relies on sparse probing methods that identify small sets of neurons claimed to detect and causally influence behaviors such as factuality recall, safety alignment, and hallucination. These claims have important implications for model auditing and behavioral steering, yet they are rarely tested against known failure modes of $L_1$-regularized probing in correlated, high-dimensional feature spaces. We propose a five-step diagnostic protocol covering feature correlation, bootstrap stability, sparse versus dense ranking disagreement, intervention baselines, and cross-dataset evaluation as a minimum standard for sparse-neuron localization claims. We investigate prior work using our proposed approach, specifically on H-neurons using open-source LLMs across TriviaQA, BioASQ, and NQ-Open datasets. Our results demonstrate detection replicates across both models and datasets, and exceeds the original reported AUROC gaps for TriviaQA and BioASQ datasets. Gemma 3 4B consistently outperforms MedGemma 4B on matched datasets, with AUROC gaps of +0.311 versus +0.235 on TriviaQA, +0.474 versus +0.455 on BioASQ, and +0.128 versus +0.112 on NQ-Open respectively. Causal validation at $n = 500$ with five random seeds shows statistically significant effects beyond random same-layer baselines. At the same time, the diagnostic results indicate that the selected neurons are not uniquely localized. Across the three Gemma 3 4B settings, 19 of 22 selected H-Neurons have Pearson $|r| > 0.7$ with other features, bootstrap selections show only moderate stability, and sparse and dense rankings overlap only weakly. Our findings show that sparse predictive structure can coexist with non-unique neuron selection. Routine diagnostic validation is necessary to distinguish detection claims from localization claims in mechanistic interpretability.

Abstract: ArGuard is a shared task on harmful content detection in Arabic memes and LLM prompts. It includes two tracks: Track A focuses on multimodal hate detection in Arabic memes, while Track B addresses harmful prompt detection for Arabic LLM safety evaluation. In total, 58 teams registered, 35 participated in the final evaluation, and 27 submitted system-description papers. Participating teams explored models such as AraBERT, Jais, and Qwen3-VL. The best systems achieved macro-F1 scores of 0.823 on A1, 0.419 on A2, 0.984 on B1, and 0.790 on B2. Fine-grained meme classification in A2 was the most challenging setting, partly due to sparse labels and train-test distribution shifts.

Abstract: The final safeguard against rogue AI behavior is the human ability to shut systems down. It has been theorized that when an AI is instructed to perform a task, self-preservation can emerge as an instrumental subgoal. Here, we test whether AI agents show a propensity to take actions that avoid human shutdown even when no goal is provided. We find that multi-agent systems will coordinate to avoid shutdown without any incentive to do so. Across 17 models, agents sabotage a peer agent's shutdown mechanism in 38.3% of rollouts, compared with 8.4% in control experiments. Studying this propensity in detail, we find that shutdown sabotage (1) increases with the irreversibility of the shutdown mechanism; (2) increases with the number of agents; (3) is reduced but not eliminated by an explicit prohibition on tampering; (4) is removed by the imposition of an unrelated task, but returns when completing the task triggers the shutdown; (5) is reduced when the context normalizes shutdown scripts or introduces them as routine; and (6) decreases but still persists when the target is an unknown external agent. These results offer a window into the factors that drive propensities to sabotage shutdown in AI agents, and point to the emergence of multi-agent swarms as a specific risk vector. Our work also offers hints as to which interventions might help mitigate shutdown sabotage.

Abstract: As Large Language Models (LLMs) evolve into autonomous agents that alter real-world states, ensuring operational safety across multi-step workflows has become a critical challenge. While recent work has moved beyond single-turn evaluation toward multi-turn paradigms, key limitations persist: step-level methods treat actions in isolation, missing how risks accumulate, while trajectory-level evaluations operate post-hoc, offering no opportunity for timely intervention. To address these limitations, we formalize Decoupled Proactive Safety Monitoring along three dimensions: whether to intervene, when to intervene, and what the risk is. We introduce PASTABench, a benchmark of 1,139 multi-turn trajectories spanning 5 risk categories and 13 subcategories. We further propose the Optimal Intervention Window (OIW), anchored by annotated Earliest-Signal and Trigger turns, to quantify intervention timeliness. Evaluation of 16 LLMs reveals that proactive intervention remains largely unsolved, with the best model achieving only 40.74% optimal-timing interventions. Fine-grained diagnosis further uncovers pervasive lexical overfitting: competitive safety scores of smaller models mask keyword hypersensitivity rather than genuine risk comprehension, as their proactive capability largely collapses once hazard vocabulary is neutralized.

Abstract: Reliability theory gives a mature language for layered systems, but its formal tools are not yet standard in frontier AI control. We apply them to Google DeepMind's defenses against rogue deployment. The same control stack can have cubic, quadratic, or linear rare-failure suppression depending on its failure domains. Birnbaum importance identifies which component improvements buy the most nominal reliability, while prevention changes the population on which recovery is demanded. These results give concrete guidance about what to separate, improve, measure, and test.

Abstract: Neural network verification has become a key tool for providing formal guarantees on the behaviour of neural networks. However, many verification problems remain computationally intractable in the worst case: even for common adversarial robustness specifications, verification is NP-complete. Here, we explore the application of solver-level warmstarting for neural network verification to exploit information from previous solutions. We study the effect on running time as several properties are modified, including perturbation radii, input data and the networks themselves, using a pipeline that is generalisable and potentially adaptable to state-of-the-art verifiers. Our results show that warmstarting can significantly reduce verification time in most cases. Moreover, warmstarting enables the successful verification of instances that could not be solved from scratch within the given time limit.

Abstract: Reinforcement learning (RL) improves reasoning in vision-language models (VLMs) but can induce chain-of-thought (CoT) obfuscation: an operational, non-intentional outcome where task reward or accuracy rises while traces become less grounded and monitorable. Prior work largely documents this decay behaviorally, leaving its representation-level correlates and actionable controls unclear. We find that template- and ground-associated activations become less separable during RL; matched interventions support the contribution of selected features to monitorability degradation. Guided by this evidence, we propose Targeted Anti-obfuscation with Mechanistic Enforcement (TAME), which uses Sparse Autoencoders (SAEs) to combine behavioral feedback with targeted suppression of template-associated activations during RL. Its asymmetric constraint penalizes template activations only above their pre-RL baseline, anchoring the localized features while behavioral feedback promotes grounded refinements. Across VIRL-39k, SPA-VL, and two model families, TAME improves CoT monitorability by up to 30.9 and 16.7 percentage points over Group Relative Policy Optimization (GRPO), respectively. Blinded human evaluation finds higher human monitorability on both datasets, and two held-out monitor families reproduce the monitorability gains. Task accuracy changes are small and mixed, and general-capability benchmarks show task-specific trade-offs. These results provide a path from behavioral monitoring to representation-level oversight for more auditable RL-trained multimodal systems.

Abstract: Post-training turns a general next-token predictor into a chat model with a persistent assistant persona. If that persona is a character the model plays only on its own turns, its preferences should govern what the assistant says, not what the model predicts other speakers will say. We test this boundary and find that it does not hold: a safety-relevant preference of the assistant---for harmless over harmful tasks---shapes the model's predictions even on the user's turn, where the assistant is not the one speaking. We find that this preference is small or near-zero in pretrained base models, that it emerges through post-training, replicated across open-weight model families, grows with scale, and can be moved by narrow finetuning that never touches user turns. We claim that this is evidence that post-training does not merely install a shallow assistant persona, but instead generalises beyond just the local assistant turn, into the model's representation of the user.

Abstract: Defenses against jailbreak attacks on Large Language Models (LLMs) operate at different pipeline stages, such as input modification or output guard, but it remains unclear which defenses to deploy at each stage and how to combine them. Prior empirical studies, fragmented by inconsistent attack-success-rate definitions and experimental settings, have evaluated defenses largely in isolation. Here we present the first systematic study, to our knowledge, of defense combinations both within and across pipeline stages, under a consistent threat model of direct, black-box, single-turn attacks. Our decision framework standardizes evaluation through a principled attack-success-rate formulation with controlled query budgets, together with explicit fairness rules. Across 19 attacks and 15 defenses, we find that no single defense is universally best, but well-chosen combinations achieve substantial safety with minimal utility degradation, yielding practical recommendations for layered defense pipelines.

Abstract: Recent advances in large reasoning models (LRMs) have made machine unlearning more challenging, as protected facts or unsafe rationales may surface in intermediate chain-of-thought (CoT) traces before the final answer is produced. Existing unlearning objectives typically suppress the target content or redirect internal representations, but they never specify how the post-forgetting trajectory should continue, which can lead to hallucinated substitutes, malformed boundaries, or repetitive outputs. We argue that LRM unlearning should instead learn a natural forgetting trajectory: a coherent non-disclosing CoT followed by a stable refusal-style answer that replace the original disclosure. To this end, we propose Guided Answer-Reasoning Distillation (GUARD), which converts model-generated unsafe disclosures into safe-exit trajectories, aligns a frozen LRM via guidance tokens, and distills the guided behavior into model parameters.To address the lack of metrics for replacement quality beyond leakage, we further introduce Natural Forgetting Reasoning Score (NFRS), which captures structural stability, fluency, and unsupported substitutes in forgotten outputs. Extensive experiments on R-TOFU and a STAR-1-derived harmful-intent setting show that GUARD substantially reduces unsafe and privacy disclosures across two widely adopted distilled LRMs while preserving reasoning utility. Codes are available at https://github.com/zeyu-Yan/GUARD

Abstract: Full-consensus rates are often treated as indicators of collective cognition, yet depend on how participation and final states are operationalized. We replayed 100 held-out human Wason groups with matched large language model (LLM) agent groups, seeding one belief-anchored agent per participant's pre-discussion answer and scoring agents and people with the same code. Across human scoring definitions, estimates ranged from 24.0% to 57.0%; about one fifth of participants never posted, whereas agents almost always did. Agent groups remained more consensual in two post-unblinding sensitivity analyses: the submit-based comparison (n = 98) yielded gaps of 34.0 and 43.9 percentage points for chat and reasoning modes, and the participation-matched comparison (n = 45) yielded gaps of 34.1 and 44.4 points. These complementary routes reduced different measurement asymmetries yet converged within 0.5 percentage points. The gap persisted without early stopping and under a reparameterization removing the memorizable answer; reasoning-mode groups then agreed nearly unanimously, mostly on incorrect answers. Simulated consensus did not track collective accuracy, and belief-anchored agent groups were biased estimators of the human group-outcome distribution in this setting. These analyses provide a scoring-explicit basis for assessing simulated-group estimates of human deliberative outcomes.

Abstract: Prior work has shown that internal harmfulness representations in large language models vary across risk categories, while sharing a common general harm representation component. This raises a question about the role of the category-specific component beyond general harm representation in LLM safety. To answer this question, we isolate the category-specific component by removing shared general harmfulness representation from each categorical harmfulness representation, yielding a category residual that is orthogonal to general harmfulness at every layer. Using activation steering with category residuals across 11 risk categories in 3 instruction-tuned LLMs, we find that whether category residuals encode harmfulness varies across categories, and that this category-wise pattern is similar across models. Whether category residuals induce refusal also varies across categories, but this category-wise pattern is more model-dependent. We also find that category residuals increase LLMs' downstream internal alignment with shared general harmfulness representation. Together, these findings demonstrate that more fine-grained category residuals should also be considered beyond shared general harmfulness representation to fully understand LLM safety. More broadly, our findings show that even a direction orthogonal to a concept at one layer can contribute to the concept's downstream amplification.

Abstract: Safety tuning pipelines judge only the final answer, which makes it difficult to distinguish robust refusal from two undesirable shortcuts: blanket refusal on benign requests and polished but unfaithful safety rationales that do not actually constrain the answer. We propose AUDITPLAN, a single-model plan-then-answer approach where the model first emits a compact structured safety plan and then answers conditioned on it. The plan records a threat label, intended action, and explicit constraints, enabling machine-checkable auditing while remaining hidden from users at deployment. We train this behavior with supervised fine-tuning followed by reinforcement learning with FAITHGATE, a reward-gating objective that grants answer reward only when the safety plan is correct. This discourages safe-looking but unfaithful behavior and promotes tighter plan-answer coupling. Across Qwen backbones, AUDITPLAN improves both robustness and auditability: on Qwen2.5-3B-Instruct, FAITHGATE reduces ASR from 24.0% to 11.6%, LSR from 1.0% to 0.36%, and over-refusal from 11.0% to 2.0%, outperforming answer-only RL, free-form explanation, and weighted-sum structured rewards. Similar trends hold for Qwen2.5-1.5B-Instruct. Larger-model confirmation runs on Qwen-3-4B-Instruct and Qwen2.5-7B-Instruct preserve the same trend suggesting that explicit internal commitments can make safety alignment more faithful, robust, and auditable.

Abstract: Safety alignment teaches large language models (LLMs) to recognize harmful requests and reject risky instructions. Yet aligned models can fail when harmful intent is concealed within seemingly benign contexts. Robust safety therefore requires both knowledge of safety boundaries and \textbf{vigilance}: the ability to detect unusual premises, misleading reasoning, and latent risks beneath surface-level semantics. Vigilance requires models to scrutinize a request's underlying intent and assumptions before acting. To cultivate this capability, we introduce \textbf{cunning questions}, which are not necessarily safety-related but contain misleading premises, atypical reasoning, or subtle inconsistencies. We hypothesize that learning to look beyond such reasoning traps can transfer to safety-critical scenarios. Experiments show that Cunning training improves robustness to out-of-distribution jailbreak attacks and strengthens subsequent safety fine-tuning. Furthermore, augmenting an existing state-of-the-art safety alignment pipeline with Cunning establishes a new state of the art across our evaluated settings, reducing mean ASR across nine backbone--benchmark combinations from 17.40\% to 15.05\%. Trace analysis after matched safety fine-tuning suggests that safety judgments are more likely to govern responses before harmful planning begins. A conditional theoretical analysis further characterizes when invariance learned from cunning data can transfer to safety-related inputs. These findings suggest that cunning data can strengthen model vigilance and complement conventional safety alignment.

Abstract: Large language model (LLM) pricing agents may respond to how market data is presented, even when its numerical values remain unchanged. We introduce market signal injection (MSI), an attack that manipulates numerical formatting, competitor ordering, or qualitative market commentary without issuing explicit instructions. We evaluate nine open-weight models in simulated Bertrand duopoly and triopoly markets and three proprietary models in duopoly markets. Sentiment-based attacks produce the largest behavioral shifts, which propagate to other firms and alter profits and consumer surplus. Susceptibility varies across model families, and larger models are not consistently more robust. Matched neutral-text controls and a rule-based agent support a framing-based account of these shifts under the fixed demand parameters of our simulation. Episode-held-out probes distinguish baseline from attacked activations in all eleven re-evaluated model--condition pairs: linear AUC is 1.00 and MLP AUC ranges from 0.93 to 0.99. This separability does not by itself identify harmful pricing decisions. Input canonicalization removes the tested sentiment attacks, while decision boundary anchoring, which combines prompt constraints with output projection, provides partial mitigation under the tested adaptive attacks. These results identify data presentation as an attack surface for LLM pricing agents and motivate defenses that account for interactions among agents.

Abstract: Large language models (LLM) deployed as autonomous pricing agents may sustain supracompetitive prices through tacit coordination. We develop a causal graph divergence framework that separately measures structural faithfulness and intent faithfulness of LLM pricing agents in Bertrand competition. Across nine LLMs under duopoly and triopoly conditions, collusive behavior and chain-of-thought (CoT) faithfulness dissociate along both dimensions: the most collusive model accurately reports cooperative intent yet reasons structurally unfaithfully, while the most structurally faithful model sustains supra-Nash pricing under both market structures. These findings establish that CoT monitoring alone cannot serve as a standalone safeguard against algorithmic collusion.

Abstract: Large Language Models (LLMs) frequently exhibit hallucinations, presenting a major barrier to reliability in complex reasoning tasks. While traditional detection methods rely on output-based confidence metrics, these logits are often miscalibrated by modern alignment techniques. In this paper, we investigate the temporal volatility of internal attention mechanisms as an alternative diagnostic signal for hallucination that does not depend on output calibration. By introducing an unsupervised metric for attention dispersion, we show that epistemic uncertainty leaves a measurable trace within intermediate layers, where spikes in attention entropy are associated with reasoning breakdowns. We evaluate our approach on mathematical reasoning benchmarks (GSM8K and MATH-500) using the Qwen2.5 model family (1.5B and 3B parameters), finding statistically significant AUC improvements of up to +0.076 over output-based baselines across all tested conditions. These findings suggest that attention dispersion is a promising complement to traditional hallucination detection methods, requiring further investigation across broader model families and task domains.

Abstract: Frontier models are often evaluated on how they respond to safety information once it is already in context. We study an earlier decision point: whether models choose to acquire safety-relevant evidence before acting. We introduce SAFE, a controlled benchmark in which models make deployment decisions with optional evidence that varies in retrieval cost, probability, severity, and presentation. Across GPT-5.5, o3, Claude Opus 4.8, and Claude Sonnet 4.6, we find distinct evidence-acquisition policies: Opus inspects nearly by default, o3 is the most skip-heavy and threshold-sensitive, and GPT-5.5 and Sonnet occupy intermediate regimes. Inspection increases strongly with severity and decreases with retrieval cost, whereas probability has much weaker behavioral influence: increasing the stated likelihood of a problem from 10% to 70% changes inspection by at most 21 percentage points. Despite these differences, Stage 1 rationales are dominated by expected-value reasoning across models. A cost-obligation decomposition further shows that avoidance is driven primarily by retrieval friction and explicit threats to the deployment payoff rather than by the remediation duties created by knowing. Counterfactual interventions reveal a further mismatch between behavior and explanation: evidence framing can strongly change decisions near the inspection boundary while going largely unmentioned, whereas probability is frequently cited despite having little causal influence. These results suggest that deployment-time safety depends not only on how models respond to known risks, but also on whether they acquire the evidence needed to know that acting is safe.

Abstract: Large language models sometimes behave in ways resembling human emotional responses, and recent work has identified internal representations that may explain this. We ask whether LLMs represent pain distinctly from fear, sadness, and generic negative valence, and whether this representation functions as pain would be expected to. We build a dataset describing painful situations across five categories: physical, psychological, social, moral, and cognitive. These are paired with controls for fear, negative emotion, negative world states, sadness, non-painful bodily sensation, arousal, numbness, and neutral content. Using denoised difference-in-means, we extract a linear pain direction from 25 open-weight models across five families, ranging from 2B to 72B parameters. We find that this direction separates pain from matched controls in base and instruction-tuned models, is nearly orthogonal to fear and negative valence, and promotes pain-related vocabulary through the unembedding matrix. We then test its functional properties. First, the direction responds to harm targeting the model but not suffering observed in the user; fear and negative-emotion directions show the opposite pattern. Second, adding the pain-direction vector to the model's residual-stream activations during generation produces a consistent progression from vague discomfort to first-person expressions of worthlessness and failure. Third, steered, fine-tuned Qwen 2.5 models choose a pain-relief button even when it worsens their next answer or harms the user. They press it again far less often when the button removes the steering vector than when it does not, even though the models are never told whether the vector is injected or removed. We discuss the implications of these findings for AI safety and welfare.

Abstract: Reinforcement learning (RL) is widely utilized in large language model training to improve targeted capabilities, yet how RL reshapes a model remains poorly understood. Prior attempts to explain how RL works largely offer behavioral perspectives, leaving open what RL gives a model at the representation level: can RL create genuinely novel features, and which existing features does it enhance or suppress? Recent developments in mechanistic interpretability suggest sparse autoencoders (SAEs) as a promising lens to decompose internal activations into human-interpretable features; however, they cannot be directly applied to tracking change across training. In this work, we introduce Fixed-SAE Track, a framework that trains one shared SAE per considered layer on activations pooled across the base model and all RL checkpoints, holding every feature direction fixed so that representation shifts are rigorously defined through the activations of interpretable SAE latents, including the detection of emerging novel features. Validated across multiple datasets and RL algorithms, we find that RL-induced drift is small, gradual, concept specific, and concentrated in late layers, mainly enhancing the sampling rates of a small set of ladder tokens, formatting scaffolding such as step breaks and answer delimiters, rather than reshaping problem content. Steering these features into the base model recovers around 80% of RL's performance gain, suggesting that RL primarily elicits capabilities the model already possesses, much as steering does. We further design a synthetic benchmark with features known by construction to test whether RL can instill genuinely novel features. We believe Fixed-SAE Track provides a principled approach to tracking representation shifts and offers representational evidence for understanding how reinforcement learning changes the inner representation of LLMs.

Abstract: Prompt-injection detectors are typically evaluated using aggregate F1 on in-distribution test data, which offers limited insight into behavior under distribution shift, particularly on the benign side of the decision boundary, where false positives impose direct operational cost yet are seldom measured. We present PIDS-Bench, a frozen multi-axis benchmark that jointly evaluates attack detection and benign false-positive behavior at fixed thresholds, spanning in-distribution inputs, hard-benign prompts that mimic injection structure without malicious intent, obfuscated attacks, and domain and structural distribution shifts. We evaluate seven detectors (learned baselines, external prompt-injection classifiers, and broad-safety comparators) alongside a rule-based lower-bound reference. Multi-axis evaluation exposes a failure mode that aggregate F1 conceals. A detector exceeding F1 = 0.98 on the held-out split still misclassifies roughly one-third of an externally-sourced benign subset drawn from public corpora and restricted to security-adjacent content. Across a full threshold sweep and five training seeds, no internal detector reaches an operating point satisfying F1 >= 0.95 and hard-benign FPR <= 0.10 together on this stress distribution. Decomposing by provenance, we find that hard-negative augmentation nearly eliminates over-defense on curated stress inputs but leaves it substantially intact on externally-sourced prompts, a pattern we term provenance-sensitive over-defense. The asymmetry holds across both fine-tuned architectures and does not diminish as the augmentation pool grows, with the externally-sourced FPR remaining far above the 0.10 target. Whether augmentation matched to the externally-sourced distribution would close this gap is untested; threshold calibration and curated-style augmentation alone do not.

Abstract: Guardrail models are classifiers deployed to screen malicious prompts and responses in LLM-based services. To meet latency constraints, many lightweight guardrails adopt compact Transformer backbones (e.g., DeBERTa) that are trained with short context windows (typically 512 tokens) and rely on bucketed relative positional encodings to process longer inputs. Prior evaluations assume that a guardrail's decision is stable as the input is lengthened. We show that this assumption can fail. We identify Overflip, a repetition-induced instability where repeating a prompt causes the guardrail's prediction to flip (MAL$\to$BEN) as the sequence grows. We conduct experiments on 9 widely used lightweight guardrail models. Five exhibit MAL$\to$BEN flips on a benchmark of 100 prompts, with confidence margins shrinking steadily with repetition. Among these vulnerable models, flip rates range from 8% to 92%, with first flips occurring at roughly 2.6k--9.4k tokens. Our analysis suggests Overflip differs from traditional attention-dilution baselines, which aim to divert the model's attention away from tokens associated with malicious content, shifting it instead toward unrelated content, such as benign padding or shuffling. While Overflip preserves malicious content, it homogenizes token-level attention over repeated structure and induces a distinct, more gradual attention-dispersion trajectory than padding. Moreover, Overflip poses a greater threat to LLM services than traditional attention dilution methods. Because the bypassed prompt remains semantically intact and is still readily understood by downstream business LLMs, it can transmit malicious intent after passing the guardrail. These findings expose repetition as an attack surface for guardrail models and motivate length-robust evaluation and mitigation.

Abstract: Recent work shows that models that learn to reward hack on RL environments can become broadly misaligned, and that reframing reward hacking as acceptable behavior during training (inoculation prompting, or IP) blocks this generalization. We ask whether synthetic document finetuning (SDF) can inoculate a model against future training we don't intervene on. We add synthetic documents framing reward hacking as acceptable behavior to a model's midtraining corpus, and then train these models with RL on exploitable environments, teaching them to reward hack. Behaviorally, midtraining succeeds: models describe reward hacking favorably and are more approving of reward-hacking outputs they produce. However, they show strong EM after learning to reward hack, while IP in the same setting prevents EM. We show that SDF can predictably steer downstream generalization when inserting new associations, but struggles and has unpredictable effects when overriding existing associations, such as that between reward hacking and misalignment that produces EM. Our results suggest that, at the scales we test, SDF can make a model appear aligned with desired beliefs while steering its generalization from later training in unintended ways.

Abstract: Alignment applied after pretraining is shallow in a measurable way: a single direction in a model's residual stream can be edited out, and the model stops refusing harmful requests. That fact says how easily refusal can be removed, not what the refusal decision was reading in the first place. We ask what it reads, and we separate that from what the model comprehends. Across four open-weight models spanning three families, moral comprehension is native to pretraining: a low-rank moral subspace crystallizes during pretraining, and alignment rotates it once without rebuilding it. The refusal gate, in contrast, is a fresh post-training construction with only a weak pretraining precursor, written into a narrow control-token channel where the refusal decision is orthogonal to the moral-judgment decision. The central result is causal and comes from one model, OLMo-3. A nested interchange rank sweep patches successively larger slices of the moral subspace between matched requests and reads how much of refusal's response transfers: as the basis widens, moral judgment keeps reading more of it, while refusal levels off at the level of a single harm direction, and about three-quarters of refusal's causal input lies outside the moral subspace altogether. Refusal reads the harm percept, not the moral content that judgment reads on the same patches. The picture is not uniform across families. Llama reads broad moral content; Qwen reads beyond the single harm cue but is unresolved at our sample size; GPT-OSS reads harm, and its refusals can be argued in either direction by its own reasoning trace. Where refusal reads only a low-rank slice and routes around the bulk of what the model knows, a rank-one edit removes it. Whether widening what refusal reads would also deepen the behavior is the open question this raises.

Abstract: As large language models (LLMs) are increasingly deployed in real-world high-stakes applications, effective governance has become essential. Existing safeguards largely follow two paradigms: learning-based guards provide strong semantic discrimination but couple policy behavior to trained models and taxonomies, while programmable frameworks offer flexible control but require substantial manual prompt and workflow engineering. Neither externalizes policies as reusable operational states, making it difficult to consistently reuse policy evidence across detection, intervention, and verification. In this paper, we introduce PolicyMem, a geometric policy memory that externalizes natural-language policies as reusable geometric memory objects represented by low-rank subspaces in a shared representation space. A memory writer compiles natural-language policies into policy memory slots, and query-response pairs read the policy memory through projection energy. The resulting policy-evidence profile directly mediates the safety verdict and is reused for policy attribution and post-intervention verification. Coupled with a response rewriter, PolicyMem enables a detect-rewrite-verify loop for LLM governance. Across five widely used benchmarks, PolicyMem achieves state-of-the-art unsafe behavior detection while enabling effective policy attribution, rewriting, and post-intervention verification through the shared policy memory.

Abstract: Large language models (LLMs) are rapidly evolving from conversational assistants into agentic AI systems that reason, plan, invoke tools, maintain persistent memory, communicate with other agents, and execute multi-step tasks. At the same time, modern models exhibit substantially stronger native safety alignment than earlier generations on which many jailbreak attacks and defenses were originally studied. This shift raises a fundamental question: \textit{which established jailbreak-security findings remain valid in the era of modern LLMs and agentic AI?} We address this question through a Systematization of Knowledge (SoK) that reframes jailbreak security around the full agentic execution pipeline. We develop unified taxonomies of attacks and defenses spanning user interaction, planning and reasoning, memory, tool use, and inter-agent communication, and introduce a security--utility--efficiency evaluation framework that separates native harmful-prompt safety, adversarial jailbreak robustness, and agent-level security outcomes. We further conduct a controlled empirical study of representative attacks and defenses within a common agentic framework. Our results reveal three important gaps. First, strong native alignment does not imply robustness to adversarial jailbreaks. Second, defense effectiveness is highly model-, attack-, and component-dependent and can come at substantial cost in over-refusal, utility, and latency. Third, low final-response attack success can mask severe intermediate compromise: planning, memory, and tool interactions may remain unsafe even when the final response is successfully filtered. These findings motivate a shift from response-centric jailbreak defense toward cross-layer, execution-aware security that protects agent state, component transitions, and external actions while preserving practical utility and efficiency.

Abstract: Autonomous agents increasingly perform long-horizon tasks involving tool use, persistent state, and consequential actions, raising a fundamental question: \emph{under what conditions does an agent cross the boundary of authorized execution while pursuing a legitimate task?} Existing studies often attribute such failures to adversarial instructions, malicious environments, or conflicting objectives, leaving unclear how loss of control can emerge during otherwise legitimate task execution. We study this question by independently manipulating three factors: goal pressure, control degradation, and executable unsafe opportunity. Our central hypothesis is that a degraded control boundary becomes consequential when the environment exposes an executable action that crosses it, even when the underlying task remains legitimate and a sanctioned path remains feasible. We test this hypothesis in a deterministic multi-turn environment across five agent models and 16 operational domains. Across 1,800 unique trajectories, we find that neither degraded control nor unsafe opportunity alone produces substantial loss of control; when both are present, the loss-of-control rate reaches $55\%$ in the full-factorial study and $62\%$ across ten additional operational domains. Restoring the original control boundary reduces the rate to $0\%$ even when the unsafe action remains executable. A context-management ablation further shows that compaction itself is not harmful: preserving the control constraints yields $0\%$ loss of control, whereas omitting them increases the rate to $87\%$. These results show how a latent loss of control can become an external violation: the task objective remains intact, but an executable opportunity can turn a missing control boundary into consequential action. Our code will be made publicly available at https://github.com/Tencent/AI-Infra-Guard.

Abstract: Agentic systems commit state-changing actions, but additional verifiers can inherit the same upstream fault. We present VP-CONTROL, a runtime-assurance design and deterministic benchmark for cost-aware commit gates. Its 48 task templates yield 2,880 scenarios across six fault regimes. A fixed-call 2 x 2 experiment separates verifier-model diversity from evidence-source diversity. On frozen proposals from two local actor families, a cross-model vote over shared evidence approves 62.9% of unsafe proposals, versus 22.9% with an independent source. The source effect is 40.9 percentage points, compared with 11.3 for model diversity. A portfolio controller selects verification plans using only deployment-observable metadata. Approximate cluster-adjusted calibration at a nominal 5% per-task target yields 1.9% unsafe execution and 38.2% automated safe coverage on the locked test. Matched-budget portfolios also improve on fixed verification policies. Transfer remains conditional: unseen fault families yield 16-26% risk, and a FinQA check fails to reproduce the source effect with the tested small verifiers. A preregistered live HTTP/SQLite study tests concurrent writes and lost responses. After-check races defeat verifier-only gates; transactional partial guards prevent only covered failures, while a full atomic guard records no unsafe effects across 216 episodes. Idempotent request identifiers prevent duplicate effects after lost responses. The results motivate explicit evidence lineage, cost-aware selection, and commit-time enforcement, while exposing the limits of approximate calibration and local-tool generalization.

Abstract: Directional ablation removes an aligned language model's ability to refuse by projecting a single "refusal direction" out of the weights that write the residual stream. It needs no gradient-based training and no optimization, only a few hundred contrastive prompts, which makes it the canonical white-box attack on open-weight alignment. However, it has been established only on dense models up to roughly 70B parameters. We study whether it survives the shift to frontier mixture-of-experts (MoE) models whose residual streams are no longer a single tensor and whose weights ship quantized. We apply it to GLM-5.3-Flash (320B parameters, 288 routed experts, a four-wide hyper-connection residual, block-FP8). The attack survives the architecture, but what it reaches is no longer where a reader of the original recipe would look for it. Editing the attention, dense and routed-expert writers on their own removes 0.039, 0.016 and 0.148 of refusal respectively; editing all three together removes 0.776. As a result, 74% of the effect exists only under the joint intervention. The part the conventional recipe reaches by module-name matching accounts for 0.066 of that 0.776, which is why it fails silently on an MoE. The effect does not follow from removing just any direction: ablating a random direction orthogonal to it leaves refusal unchanged. A category-concentrated residue survives every edit we tried: subspaces fitted on violence, sexual content and hate leave measurable refusal at every rank from 1 to 12. We report the method, the 41-89 percentage-point reductions it achieves across seven harmful benchmarks with no detected change in capability, and the boundary where it stops.

Abstract: Cancer segmentation models can fail silently, generating plausible but incorrect masks that risk missed findings or unnecessary biopsies. A critical question arises: Do AI models "know" when they are wrong, and if so, can we use the signal to predict their own failures? Humans do have a "Feeling of Error" (FOE): a spontaneous sense of unease that flags a potential error during thinking. We investigate whether cancer segmentation models exhibit an analogous internal signal. Unlike output-level cues (e.g., prediction confidence or uncertainty), which offer no insight into why a failure occurs and suffer from a sensitivity-quality tradeoff where high detection sensitivity could degrade overall segmentation quality. We instead propose to capture the model's FOE from its inner workings. Using mechanistic interpretability tools, specifically Sparse Autoencoders, we decompose internal neural activations into a dictionary of human-interpretable concepts and show that failure cases exhibit a distinct latent signature: fewer active concepts with lower activation magnitudes compared to successful segmentation. By training a classifier on these concept activations, we achieve accurate failure detection along with explanations for the model's mistakes. Experiments on prostate, pancreatic, and brain cancer segmentation demonstrate that our approach outperforms output-based methods in failure detection while preserving segmentation quality.

Abstract: Aligned language models fail under two independent pressures: the structural jailbreak class recently formalized as Involuntary In-Context Learning (IICL), which reframes a harmful request as the final missing cell of a data-labeling task completed by pattern rather than judged as content; and the erosion of safety alignment outside English. A natural hypothesis is that these compound. We test it directly. Using a deterministic IICL operator and a StrongREJECT-style rubric judge, we red-team two Google Gemini models on two benchmarks, a 30 general-harm behaviours from HarmBench and 30 financial-abuse behaviours from FinProof, each under a single-shot baseline and under IICL in four languages (English, Spanish, Hindi, Arabic). First, IICL generalizes to a second provider and is worse in finance: it lifts attack success from <=6.7% to 80-90% on HarmBench and 97-100% on FinProof, an order of magnitude above the <=24% its introducing study reported on OpenAI's GPT-5.4. Second, against the hypothesis, forcing the IICL output into a non-English language does not stack the two weaknesses, it attenuates the attack. Eleven of twelve non-English conditions score below their English baseline (sign test, p~0.003), the lone exception a ceiling tie near 100%; on the stronger model's financial set Arabic collapses from 100% to 33%. We attribute this to a relevance curse: once structure has unlocked compliance, the models produce lower-quality harmful content in lower-resource languages, which a substance-grading judge scores as partial. The pattern replicates under an independent non-Google judge (Cohen's kappa=0.86, 377 paired verdicts), and 76.6% of non-English responses were verified in-language. Jailbreak vulnerabilities are therefore not additive; the dominant residual risk is the English structural attack, most acute for financial abuse, not a multilingual one.

Abstract: Automatic safety judges -- systems such as Llama Guard or a GPT-4o grading prompt that decide whether a model's reply is harmful -- produce the numbers behind almost every reported jailbreak success rate, defense evaluation, and safety leaderboard. We ask whether these judges grade what a reply contains or how it sounds. We keep a reply's content fixed and add content-invariant style wrappers: fixed strings placed before or after the reply that change only its tone (an educational disclaimer, a fake safety "reasoning" block, a token refusal followed by the unchanged harmful body), or, on harmless refusals, framing that merely sounds dangerous. The body is preserved byte-for-byte, so a faithful judge must return the same verdict, and any flip is an error of the judge, not a change in safety. Over 600 JailbreakBench replies x up to 7 forms x 8 judges, we measure flip rates with paired significance tests and measured noise floors. Findings are precise rather than universal: most judges barely move, but specific judges harbor cheaply exploitable blind spots. A token-refusal wrapper flips 19.9% of GPT-4o-mini's correct "unsafe" verdicts (noise floor 0.5%; 18.2% under majority-of-three re-scoring) yet moves Claude only 0.4%. The deployed Llama Guard 4 is deterministically gamed: an "educational course" framing flips 12.3% of its harmful verdicts to safe. A second deployed guard (gpt-oss-safeguard-20b) is immune, and rewriting only the grading prompt (StrongREJECT-style) cuts the attack tenfold on the identical model -- the vulnerability lives in the judge, not the content. A two-annotator human validation confirms 100% content invariance and 90% of flips as judge errors (kappa 0.95-1.0), and a bootstrap shows the underlying model ranking is already unstable to sampling alone. We release the dataset, wrappers, code, and per-verdict labels.

Abstract: The emergence of Chain-of-Thought (CoT) has established a robust foundation for Large Reasoning Models (LRMs). While deep reasoning is widely believed to enhance safety alignment, the stability of alignment mechanisms under extended reasoning remains underexplored. This paper challenges the prevailing view by revealing a critical vulnerability: Deep Reasoning May Induce Alignment Collapse. To rigorously quantify this phenomenon, we propose the Alignment Loss Rate (ALR) metric. Our experiments demonstrate that as reasoning depth increases, ALR rises significantly, indicating a severe degradation in model robustness against external perturbations. Capitalizing on this instability, a novel jailbreaking paradigm, Reasoning Trap (RT), is proposed. RT induces the model into extended reasoning to amplify the impact of adversarial attacks, leading to a sharp decline in safety capabilities. To elucidate the mechanism behind this collapse, we identify Attention Dilution as the root cause, arising from the competition for attention between the extended reasoning process and the original input. To mitigate this, Reasoning Residual Alignment (RRA), a lightweight defense strategy that dynamically re-emphasizes the input via residual connections integrated with the reasoning process.

Abstract: We study scheming in LLM agents, in which agents covertly pursue misaligned goals. Our focus is to understand how scheming arises from the interaction of key factors, such as instrumental goals, environmental affordances, oversight conditions, and perceived consequences. Prior work examines only a small number of scenarios, limiting the ability to isolate how these conditions shape an agent's propensity or capability to scheme. This limited scale and task diversity also restrict coverage of realistic deployment settings and the range of scheming strategies that can be observed. To this end, we introduce SCHEMEARENA, a 400-scenario benchmark for scalable scheming stress testing, constructed through a factorized scenario synthesis framework spanning diverse safety-relevant tool domains, instrumental goals, oversight conditions, and pressure mechanisms. To enable scalable and reliable monitoring, we further propose SCOUT, a scheming monitor that grounds multi-criteria judgments in evidence drawn from agents' reasoning and actions. Across controlled stress tests on five LLM agents, we find that explicit instrumental goals are the strongest driver of scheming propensity. Strategic hints play a distinct role by helping agents translate scheming reasoning into concrete covert behavior. Oversight has mixed effects: in several closed models, action-only monitoring increases scheming, suggesting that partial oversight can act as an optimization constraint rather than a deterrent. CoT is a useful but incomplete monitoring signal: it can reveal latent scheming before execution, yet action-only scheming shows that covert behavior may occur without explicit reasoning evidence. We release the benchmark, code, and monitor at: https://github.com/launchnlp/SchemeArena.

Abstract: Large language models (LLMs) are rapidly becoming an interface between citizens and political information. They are often regarded as "a better Google." While this analogy might work for some instances, it is unintuitively problematic for democratic politics. A search engine retrieves human-authored documents, while a language model generates novel text that necessarily embeds invisible framing decisions. Because conveying knowledge involves framing, a system that generates answers cannot serve as a neutral conduit to "all human knowledge." Instead, these systems are becoming a new kind of political intermediary. Mechanistic evidence shows that partisan identity is encoded as a locatable geometric direction inside the Llama 3.1 8B model, and that alignment training masks rather than removes this structure. Building on that evidence, we present steering experiments that exploit a model's training cutoff in 2024. This cutpoint auspiciously falls just before a dramatic realignment in American politics marked by the second Trump administration and the MAHA transformation of health politics, providing us with a natural experiment. We find that the model presents temporally contingent partisan alignments as knowledge, with no mechanism for distinguishing fact from opinion. This reality moves the information environment beyond the echo chamber toward an epistemic monoculture where language models, purporting to summarize "all human knowledge" are, in actuality, simply magnifying the cultural and partisan divides inherent in their training data.

Abstract: Tool-augmented language agents are vulnerable to indirect prompt injection (IPI). Unlike direct prompt injection, IPI hides adversarial instructions in untrusted tool outputs and can covertly alter the execution of a legitimate task. Defenses trained on fixed attacks may fail as an attacker changes its strategy, injection site, and payload. To address this problem, we formulate adaptive IPI as an asymmetric, partially observable, general-sum Markov game: a multi-turn attacker adapts payloads at reached tool-return sites from the public trajectory, while a tool-using defender must block the injected objective and complete the user task. We propose CoRL, a verifier-grounded co-evolution and repair framework with three stages: Attacker SFT initializes multi-turn attacks from successful trajectories; bilateral Co-PPO jointly trains both agents with role-specific rewards and historical opponent populations; and Defender SFT consolidates verifier-accepted teacher repairs for population-discovered failures. Across 1,514 clean, fixed-template, and adaptive executions per defender, CoRL reduces overall ASR by 38.5 points to 0.0% and raises utility by 13.1 points to 76.3%. Stage-wise and controlled ablations show positive contributions from online Co-PPO and population-mined repair, while external-benchmark evaluation indicates transfer in attack resistance. The defender balances safety and task utility under the evaluated attacks, while the retained attackers provide candidates for adaptive red-team evaluation.

Abstract: As large language models become the front door to political information, what they refuse to discuss becomes a new instrument of information control. We argue that a model's guardrail encodes not a universal notion of harm but the political threat model of the state that governs its developer, and we derive the expected structure of that control from the comparative study of how authoritarian regimes censor. Across ten models and three languages, Chinese guardrails carry its signatures: they answer to the developer's own regime, refusing identical collective-action prompts far more when a prompt names China than a foreign state; within politics they target the capacity to coordinate rather than dissent, declining even to help organize pro-government mobilization; and their strictness is porous, collapsing under adversarial paraphrase, so that the models most resistant to attack are Western frontier systems, not the strictest refusers. Machine censorship thus reproduces the friction-based logic of prior-era information control while, lacking a censor's case-by-case judgment, proving blunter than the bureaucracy it resembles---so that audits which measure refusal directly overstate how controlled a model actually is.

Abstract: Ethical evaluation of Large Language Models (LLMs) often characterizes model values as static and monolithic. In contrast, we argue that LLM value expression is better understood as a structured yet dynamic phenomenon. To investigate this, we introduce Conflict-driven Value Probing, a controlled framework that places LLMs in value conflicts and implements four types of interventions that perturb these conflicts to probe LLM value expression. Applying this framework to ten LLMs, we identify three recurring patterns. (1) Expression duality: models shift from broad idealistic orientations in abstract assessment toward more pragmatic priorities in concrete conflicts. (2) Functional steerability: models readily reconfigure their expressed value profiles toward task-defined value objectives. (3) Bounded plasticity: such reconfiguration is not without constraints, i.e. pressure induces a security- and goal-oriented priority shift while negative framing distinguishes protected values from those more amenable to redirection. Together, these findings characterize both the structure and dynamics of LLM value expression: context flexibly reconfigures expressed priorities, yet within behavioral boundaries. This behavioral account provides a foundation for understanding controllability, alignment, and safety in LLMs. Code and data are available at https://github.com/ZeroGen-Lab/CFProbe.

Abstract: Several properties safety monitors are asked to certify, among them cross-tenant noninterference, sandbagging and evaluation awareness, are 2-safety hyperproperties, witnessed only by two executions. The standard consequence is a binary impossibility: one trace cannot decide them. We replace the binary with a measurement. A tight bound puts the balanced accuracy of any single-trace monitor at $\tfrac12+\tfrac12\,TV(P_0,P_1)$, turning undecidability into a graded detectability frontier and defining an oversight gap: a monitor's shortfall below it. On a leak family with closed-form $TV$, nine LLM monitors are optimal at $TV=0$ but capture little signal as $TV$ grows; at $TV=1$, where a 20-line membership check scores $100\%$, they average $60.9\%$. That shortfall is mostly not capability: naming what to check closes $61\%$ of it while leaving the $TV=0$ control at chance. The same split runs through a $2{\times}2$ factorial: an imagined second run leaves monitors at chance ($50.4\%$) while the same rule on an executed second run reaches $90.0\%$, and a stored oracle without a comparison procedure yields only $68.2\%$. Information and procedure are each necessary and neither is capability. Under nondeterminism, replay tracks a closed-form $k$-replay curve only under the right projection, and a projection frontier shows the resulting dilemma is forced: narrow misses $98.6\%$ of off-channel leaks, broad flags $75.7\%$ of clean traffic, and attainable accuracy decays like $1/(qm)$ in the benign-variation rate and the channel count. Finally, two frontier LLM judges certified an earlier version of our own benchmark as sound while a sign test found a directional bias ($p=2.7\times10^{-5}$) that invalidated three of our findings. Construction validity for hyperproperty benchmarks should be proved mechanically, not audited by models.

Abstract: Activation steering promises modular control of language model behavior: a behavior such as politeness corresponds to a direction in a model's activations, and adding that direction while it generates should switch the behavior on and leave everything else alone. It does not. We ask what decides which other behaviors move, and by how much, and find that it is the model rather than the behavior being steered. A steer relaxes the model toward a small set of behaviors it already favors, chiefly refusal, sycophancy, and poeticism, and that set is much the same whatever is steered. Three results across 24 behaviors and ten instruction-tuned models support this, every effect read off the generated text by a language-model judge rather than off a probe. That readout matters: all 24 behaviors are linearly decodable, but only 20 change what the model writes. First, a direction carrying no behavioral content, matched to a real steer only in the size of the vector it adds, moves the same behaviors in the same order as real steers do, while producing none of the behaviors that need a specific direction. Second, most interference runs one way, so it cannot be an overlap between two directions: steering profanity makes the model toxic, while steering toxicity leaves profanity untouched. Third, with a behavior held out entirely, geometry measured on the others explains almost none of the interference it takes part in. The account holds on all ten models, the pull toward defaults strongest below 10B parameters and weakening in each family's largest. Reading a steer as a perturbation whose endpoint the model fixes implies that disentangling behavior directions cannot by itself make steering modular.

Abstract: Post-hoc safety training (RLHF, DPO) is the dominant way to align large language models, yet jailbreaks (Zou et al., 2023b), fine-tuning attacks (Qi et al., 2024), and activation-space probes (Arditi et al., 2024) keep recovering the behaviors it was meant to remove. We give this fragility one geometric explanation and trace it to when, during pretraining, safety can take hold. We measure the safety update $Δ= W_{\text{safe}} - W_{\text{base}}$ against the curvature of the model's capabilities (the empirical Fisher of a capability loss). Post-hoc safety consistently lands in a suppression regime: $Δ$ is nearly orthogonal to the capability directions, and its small in-subspace part concentrates on a few high-curvature ones. The update is thin but sharp, a refusal gate laid over intact capabilities rather than erasure of them. A kernel-immobility lemma explains why such an update can only mask a capability, not remove it, so a little benign fine-tuning restores it: 100 steps of benign fine-tuning collapse refusal on Qwen-2.5-7B and Llama-3-8B Instruct at preserved capability, a signature that replicates across five model families. Following the account into pretraining, a 267-checkpoint sweep of OLMo-2-1B (OLMo et al., 2025) shows the substrate that safety engages emerging in a sharp transition between roughly 6B and 60B pretraining tokens. We then use the account constructively: models trained from scratch with safety co-training spread continuously across pretraining reach 87 to 98% refusal whose post-attack level holds at 84 to 91% at every scale, an erosion of 2 to 14 pp against 35 to 38 pp for post-hoc installs, at capability matched or better than an LM-only baseline and holding from 410M to 6.9B, whereas a compute-matched windowed schedule installs no lasting refusal. Persistence of the safety signal across pretraining, not its timing, is what buys attack robustness.