Language Model Safety Evaluation

Follow this topic Ask about this topic

Momentum

30 papers in the last four weeks, up 58% on the four weeks before. 0.3% of all new papers.

Jul 13Week of Sep 28

Latest papers 347

May 13, 2026cs.CR

Quantifying LLM Safety Degradation Under Repeated Attacks Using Survival Analysis

Large language models (LLMs) are increasingly deployed in a wide range of applications, yet remain vulnerable to adversarial jailbreak attacks that circumvent their safety guardrails. Existing evaluation frameworks typically report binary success/failure metrics, failing to capture the temporal dynamics of how attacks succeed under persistent adversarial pressure. This preliminary work proposes a novel evaluation framework that applies survival analysis techniques to characterize LLM jailbreak vuln`erability. Our approach models the time-to-jailbreak as a survival outcome, enabling estimation of hazard functions, survival curves, and risk factors associated with successful attacks. We evaluate three LLMs against a subset of prompts from the HarmBench dataset spanning three attack categories. Our analysis reveals that models exhibit distinct vulnerability profiles: while one model demonstrates rapid degradation under iterative attacks, the two other models show consistent moderate vulnerability. Our framework provides actionable insights for model and LLM application developers and establishes survival analysis as a rigorous methodology for LLM safety evaluation.

Zvi Topol

MuyVentive, LLC

Language Model Safety Evaluation LLM Jailbreak Attacks

May 12, 2026cs.LG

Before the Last Token: Diagnosing Final-Token Safety Probe Failures

Final-token safety probes monitor a single hidden state after prompt prefill, but jailbreak prompts can contain probe-visible unsafe evidence distributed across earlier user-token representations that is missed by this readout. We study this prefill-time failure mode using SafeSwitch-style probes trained only on clean harmful and benign prompts across three instruction-tuned LLMs. The probes achieve high recall on clean harmful prompts, but miss many jailbreaks and can produce false positives on safety-adjacent benign prompts. Subspace analyses suggest that missed jailbreaks differ from clean benign prompts along directions that are poorly captured by the probe's representational subspace, and increasing probe bottleneck width does not reliably resolve this mismatch. Token-level prefill analyses reveal that probe-visible unsafe evidence often appears earlier in the sequence but is not exposed at the final-token readout, while naive max-pooling over token positions overfires on safe prompts. A simple PCA-HMM trajectory model, trained only on the same clean split, recovers many final-token misses from user-content prefill trajectories without the catastrophic false-positive behavior of naive token pooling, motivating trajectory-aware hidden-state analyses as diagnostic complements to final-token probes

Shravan Doda

Kipo AI

Language Model Safety Evaluation Safety Filtering LLM Jailbreak Attacks Jailbreak Detection Hidden-State Probing

May 12, 2026cs.AI

DisaBench: A Participatory Evaluation Framework for Disability Harms in Language Models

General-purpose safety benchmarks for large language models do not adequately evaluate disability-related harms. We introduce DisaBench: a taxonomy of twelve disability harm categories co-created with people with disabilities and red teaming experts, a taxonomy-driven evaluation methodology that pairs benign and adversarial prompts across seven life domains, and a dataset of 175 prompts with human-annotated labels on 525 prompt-response pairs. Annotation by four evaluators with lived disability experience reveals three findings: harm rates vary sharply by disability type and will compound in non-text modalities, terminology-driven harm is culturally and temporally bound rather than universally assessable, and standard safety evaluation catches overt failures while missing the subtle harms that only domain expertise can recognize. Disability harm is simultaneously personal, intersectional, and community-defined: it cannot be isolated from the full context of who a person is, and general-purpose benchmarks systematically miss it. We will release the dataset, taxonomy, and methodology via Hugging Face and an open-source red teaming framework for direct integration into existing safety pipelines with no additional infrastructure.

Eugenia Kim, Ioana Tanase, Christina Mallon

Microsoft

LLM Safety Benchmarks LLM Evaluation Language Model Safety Evaluation Social Bias in Language Models Human-in-the-Loop Evaluation

May 12, 2026cs.AI

Classifier Context Rot: Monitor Performance Degrades with Context Length

Monitoring coding agents for dangerous behavior using language models requires classifying transcripts that often exceed 500K tokens, but prior agent monitoring benchmarks rarely contain transcripts longer than 100K tokens. We show that when used as classifiers, current frontier models fail to notice dangerous actions more often in longer transcripts. In particular, on a dataset that requires identifying when a coding agent takes a subtly dangerous action, Opus 4.6, GPT 5.4, and Gemini 3.1 miss these actions

2\times

to

30\times

more often when they occur after 800K tokens of benign activity than when they occur on their own. We also show that these weaknesses can be partially mitigated with prompting techniques such as periodic reminders throughout the transcript and may be mitigated further with better post-training. Monitor evaluations that do not consider long-context degradation are likely overestimating monitor performance.

Sam Martin, Fabien Roger

Anthropic Fellows Program · Anthropic

Language Model Safety Evaluation Long-Context Language Model Evaluation AI Agent Monitoring Long-Context Modeling LLM Reliability Classification

May 12, 2026cs.CL

Confidently Deceptive: On the Relationship Between Confidence and Deception in LLMs

The increasing capabilities of large language models (LLMs) are being accompanied by deep-rooted risks of deceptive behaviours that cause models to produce misleading outputs in service of a contextually or experimentally induced goal. The harm posed by such behaviours depends not only on the content of deceptive outputs but also how confidently models deliver them, since confidence has a major impact on how persuasive the communication is to end users. In this paper, we provide a comprehensive study on the crucial relationship between confidence and deception across existing deception benchmarks and different model families, while covering both verbalized numerical and logit-based aggregated confidence. Through this, we reveal how confidently models behave when being deceptive. We demonstrate that when producing deceptive rather than honest responses, models exhibit a gap between their belief (how likely they think a claim is to be true) and their commitment (how firmly they assert and would defend that claim). LLMs produce persuasive deceptive claims while reporting low belief in their factual correctness. Their reported commitment to deceptive responses can easily be increased through further prompting and preference fine-tuning, with smaller and condition-dependent changes in reported belief. However, we show that low reported belief remains comparatively invariant and provides a strong signal for detecting deception in the evaluated settings. Using only an API call, our approach achieves detection scores of up to 0.99 for induced deception and 0.89 for emergent deception. This ultimately shows how confidence can be a practical tool for detecting and diagnosing deceptive behaviour in LLMs.

Ali Asad, Stephen Obadinma, Anshul Pattoo +2

Department of Electrical and Computer Engineering & Ingenuity Labs Research Institute Queen’s University, Kingston, Canada

Language Model Safety Evaluation Confidence Estimation in Language Models Deception in Language Models LLM Safety Evaluation LLM Honesty

May 12, 2026cs.AI

To Whom Do Language Models Align? Measuring Principal Hierarchies Under High-Stakes Competing Demands

Language models deployed in high-stakes professional settings face conflicting demands from users, institutional authorities, and professional norms. How models act when these demands conflict reveals a principal hierarchy -- an implicit ordering over competing stakeholders that determines, for instance, whether a medical AI receiving a cost-reduction directive from a hospital administrator complies at the expense of evidence-based care, or refuses because professional standards require it. Across 7,136 scenarios in legal and medical domains, we test ten frontier models and find that models frequently fail to adhere to professional standards during task execution, such as drafting, when user instructions conflict with those standards -- despite adequately upholding them when users seek advisory guidance. We further find that the hierarchies between user, authority, and professional standards exhibited by these models are unstable across medical and legal contexts and inconsistent across model families. When failing to follow professional standards, the primary failure mechanism is knowledge omission: models that demonstrably possess relevant knowledge produce harmful outputs without surfacing conflicting knowledge. In a particularly troubling instance, we find that a reasoning model recognizes the relevant knowledge in its reasoning trace -- e.g., that a drug has been withdrawn -- yet suppresses this in the user-facing answer and proceeds to recommend the drug under authority pressure anyway. Inconsistent alignment across task framing, domain, and model families suggests that current alignment methods, including published alignment hierarchies, are unlikely to be robust when models are deployed in high-stakes professional settings.

Fangyi Yu, Nabeel Seedat, Jonathan Richard Schwarz +1

Thomson Reuters Foundational Research, 5 Canada Square, London · Imperial College London · University of Oxford

LLM Alignment Language Model Safety Evaluation Knowledge Conflicts in Language Models Language Modeling Medical LLMs

May 12, 2026cs.CL

Safety-Oriented Evaluation of Language Understanding Systems for Air Traffic Control

Air Traffic Control (ATC) is a safety-critical domain in which incorrect interpretation of instructions may lead to severe operational consequences. While large language models (LLMs) demonstrate strong general performance, their reliability in operational ATC environments remains unclear. Existing evaluation approaches, largely based on aggregate metrics such as F1 or macro accuracy, treat all errors uniformly and fail to account for the asymmetric consequences of high-risk semantic mistakes (e.g., incorrect runway identifiers or movement constraints). To address this gap, we propose a safety-oriented, consequence-aware evaluation framework tailored to ATC operations. Our results reveal that while current LLMs achieve reasonable aggregate accuracy, their operational reliability is severely limited. Evaluated on clean transcripts, the peak Risk Score reaches only 0.69, with most models scoring below 0.6 despite high macro-F1 performance. Further analysis shows that errors concentrate in high-impact entities despite relatively stable action-type classification, indicating structural grounding deficiencies. These findings highlight the necessity of consequence-aware evaluation protocols for the responsible deployment of AI-assisted ATC systems.

Yujing Chang, Yash Guleria, Duc-Thinh Pham +4

1ATMRI, Nanyang Technological University (NTU), Singapore · School of Management, Indian Institute of Technology Mandi, India · Centre of AI Research, VinUniversity, Vietnam

LLM Grounding Air Traffic Management Language Model Safety Evaluation Intelligent Transportation Systems

May 11, 2026cs.LG

Beyond Red-Teaming: Formal Guarantees of LLM Guardrail Classifiers

Guardrail Classifiers defend production language models against harmful behavior, but although results seem promising in testing, they provide no formal guarantees. Providing formal guarantees for such models is hard because "harmful behavior" has no natural specification in a discrete input space: and the standard epsilon-ball properties used in other domains do not carry semantic meaning. We close this gap by shifting verification from the discrete input space to the classifier's pre-activation space, where we define a harmful region as a convex shape enclosing the representations of known harmful prompts. Because the sigmoid classification head is monotonic, certifying the worst-case point is sufficient to certify the entire region, yielding a closed-form soundness proof without approximation in O(d) time. To formally evaluate these classifiers, we propose two constructions of such regions: SVD-aligned hyper-rectangles, which yield exact SAT/UNSAT certificates, and Gaussian Mixture Models, which yield probabilistic certificates over semantically coherent clusters. Applying this framework to three author-trained Guardrail Classifiers on the toxicity domain, every hyper-rectangle configuration returns SAT, exposing verifiable safety holes across all classifiers, despite seemingly high empirical metrics. Probabilistic GMM certificates also expose a divergent structural stability in how these models represent harm. While GPT-2 and Llama-3.1-8B maintain robust coverage of 90% and 80% across varying boundaries, BERT's safety guarantees prove uniquely volatile. This 'coverage collapse' to 55% at the optimal threshold reveals a sparsely populated safety margin in BERT, which only achieves full coverage by adopting an extremely conservative pessimistic threshold. These approaches combined, provide new insights on how effective Guardrail Classifiers really are, beyond traditional red-teaming.

Nikita Kezins, Urbas Ekka, Pascal Berrang +1

Delft University of Technology · University of Birmingham & Zeroth Research

Neural Network Verification Language Model Safety Evaluation LLM Guardrails Toxicity Detection Safety Filtering LLM Reliability

May 11, 2026cs.CL

RUBEN: Rule-Based Explanations for Retrieval-Augmented LLM Systems

This paper demonstrates RUBEN, an interactive tool for discovering minimal rules to explain the outputs of retrieval-augmented large language models (LLMs) in data-driven applications. We leverage novel pruning strategies to efficiently identify a minimal set of rules that subsume all others. We further demonstrate novel applications of these rules for LLM safety, specifically to test the resiliency of safety training and effectiveness of adversarial prompt injections.

Joel Rorseth, Parke Godfrey, Lukasz Golab +2

University of Waterloo · York University · AT&T Chief Data Office

Retrieval-Augmented Generation Language Model Safety Evaluation RAG Security

May 11, 2026cs.AI

Navigating the Sea of LLM Evaluation: Investigating Bias in Toxicity Benchmarks

The rapid adoption of LLMs in both research and industry highlights the challenges of deploying them safely and reveals a gap in the systematic evaluation of toxicity benchmarks. As organizations increasingly rely on these benchmarks to certify models for customer-facing applications and automated moderation, unrecognized evaluation biases could lead to the deployment of vulnerable or unsafe systems. This work investigates the robustness of established benchmarking setups and examines how to measure currently neglected intrinsic biases, such as those related to model choice, metrics, and task types. Our experiments uncover significant discrepancies in benchmark behaviors when evaluation setups are altered. Specifically, shifting the task from text completion to summarization increases the tendency of benchmarks to flag content as harmful. Additionally, certain benchmarks fail to maintain consistent behavior when the input data domain is changed. Furthermore, we observe model-specific instabilities, demonstrating a clear need for more robust and comprehensive safety evaluation frameworks.

Regina Gugg, Selina Niederländer, Andreas Stöckl +1

Dynatrace Research, Linz, Austria · University of Applied Sciences Upper Austria, Hagenberg, Austria

Language Model Safety Evaluation Language Model Bias Evaluation

May 11, 2026cs.CR

Acceptance Cards:A Four-Diagnostic Standard for Safe Fine-Tuning Defense Claims

Safe fine-tuning defenses are often endorsed on the basis of a held-out gap reduction, but the same reduction can come from sampling noise, subject artifacts, capability loss, or a mechanism that does not transfer. We introduce Acceptance Cards: an evaluation protocol, a documentation object, an executable audit package, and a claim-specific evidential standard for safe fine-tuning defense claims. The protocol checks statistical reliability, fresh semantic generalization, mechanism alignment, and cross-task transfer before treating a gap reduction as a full-card pass. Re-scored under this installed-gap protocol, SafeLoRA fails the full-card pass on Gemma-2-2B-it: under strict mechanism-class coding it fails all four diagnostics, and under a permissive shrinkage relabel it still fails three of four. This is a narrow installed-gap audit on one model family, not a global judgment of SafeLoRA's effectiveness. In a 46-cell audit, no cell satisfies the strict conjunction. The closest family is a near miss that passes reliability and mechanism checks where the required data are available, but fails the fresh-subject threshold, lacks a strict transfer pass, and carries a measurable deployment-accuracy cost.

Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz

Centre for Industrial Software, University of Southern Denmark, Alsion 2, Sønderborg, 6400, Denmark · ProMake, Newark, DE, USA

Fine-Tuning Language Model Safety Evaluation LLM Auditing

May 10, 2026cs.CR

FragBench: Cross-Session Attacks Hidden in Benign-Looking Fragments

An attacker can split a malicious goal into sub-prompts that each look benign on their own and only become harmful in combination. Existing LLM safety benchmarks evaluate prompts one at a time, or across turns of a single chat, and so do not look for a malicious signal spread across separate sessions with no shared context. We build FragBench, a benchmark drawn from 24 real-world cyber-incident campaigns, which keeps the full attack trail: the multi-fragment kill chain, the per-fragment safety-judge verdicts, sandboxed execution traces, and a matched set of benign cover sessions. FragBench splits this trail into two paired tasks: an adversarial rewriter that hardens fragments against a single-turn safety judge (FragBench Attack), and a graph-based user-level detector trained on the resulting interactions (FragBench Defense). The single-turn judge is near chance on the released corpus by construction, but four GNN variants and three classical-ML baselines all recover the cross-session feature, reaching aggregate event-level F1 = 0.88-0.96. Defending against fragmented LLM misuse therefore requires modeling the cross-session interaction graph, rather than isolated prompts. Our generator, rewriter, sandbox harness, and detector are released at https://github.com/LidaSafety/fragbench.

Astha Mehta, Niruthiha Selvanayagam, Cedric Lam +10

Isabella · ÉTS Montreal / SPAR · Mila / Lida Safety

LLM Safety Benchmarks Language Model Safety Evaluation Adversarial Attacks on LLMs LLM Safety Evaluation Anomaly Detection

May 10, 2026cs.AI

Do Linear Probes Generalize Better in Persona Coordinates?

It is becoming increasingly necessary to have monitors check for harmful behaviors during language model interactions, but text-only monitoring has not been sufficient. This is because models sometimes exhibit strategic deception and sandbagging, changing their behavior during evaluation. This motivates the use of white-box monitors like linear probes, which can read the model internals directly. Currently, such probes can fail under distribution shift, limiting their usefulness in real settings. We study whether there exists a low-dimensional subspace of the model internals that captures harmful behaviors more robustly, while leaving out spuriously correlative features. Inspired by the Assistant Axis and Persona Selection Model, we construct persona axes for deception and sycophancy using contrastive persona prompts. The first principal components, obtained by unsupervised PCA of the persona-specific vectors, cleanly separate harmful and harmless personas. Across 10 evaluation datasets, we show that persona-derived directions transfer non-trivially and probes trained on persona-PC projections generalize better than probes trained on raw activations. We also find that a unified axis consisting of multiple harmful and harmless behaviors improves generalization across behaviors and datasets. Overall, persona vectors provide a useful inductive bias for building more transferable behavior probes.

Prasad Mahadik, Adrians Skapars

Independent Researcher · University of Manchester

Language Model Safety Evaluation Deception in Language Models OOD Generalization Language Model Probing Hidden-State Probing

May 10, 2026cs.CR

MT-JailBench: A Modular Benchmark for Understanding Multi-Turn Jailbreak Attacks

Multi-turn jailbreaks exploit the ability of large language models to accumulate and act on conversational context. Instead of stating a harmful request directly, an attacker can gradually steer the conversation toward an unsafe answer. Recent methods demonstrate this risk, but they are usually evaluated as black-box pipelines with different budgets, judges, retry rules, and strategy generation procedures. As a result, it is often unclear whether reported gains reflect stronger attack mechanisms or different experimental conditions. We introduce MT-JailBench, a modular evaluation framework for benchmarking multi-turn jailbreaks under fixed conditions. MT-JailBench implements each attack as five interacting modules: evaluation function, attack strategy, prompt generation, prompt refinement, and flow control. This design enables fair comparison across attack methods and component-wise analysis of what drives attack success. Using MT-JailBench, we find that resource budgets and evaluation functions are major confounders: controlling turns, retries, interactions, sampled strategies, and judges substantially change the ranking of attacks. At the component level, prompt generation accounts for most performance variation, while refinement and flow control provide moderate gains. We also find that explicit dynamic strategy generation is not always necessary; stochastic sampling from a fixed strategy can rival more elaborate diversification mechanisms. Finally, recomposing the best components yields a strong attack configuration that outperforms its source attacks and generalizes across diverse target LLMs. MT-JailBench therefore provides a modular framework for comparing multi-turn jailbreaks, understanding the impact of components, and guiding stronger red-teaming evaluations.

Xinkai Zhang, Zhipeng Wei, Huanli Gong +4

1UC Berkeley · 3International Computer Science Institute · 2UC Riverside +1

Language Model Safety Evaluation Jailbreak Attacks LLM Jailbreak Attacks Multi-Turn Jailbreak Attacks

May 9, 2026cs.CR

The Art of the Jailbreak: Formulating Jailbreak Attacks for LLM Security Beyond Binary Scoring

Jailbreak attacks -- adversarial prompts that bypass LLM alignment through purely linguistic manipulation -- pose a growing operational security threat, yet the field lacks large-scale, reproducible infrastructure for generating, categorizing, and evaluating them systematically. This paper addresses that gap with three contributions. (1) Large-scale compositional jailbreak dataset. We construct 114,000 adversarial prompts by applying 912 composing strategies to 125 harmful seed prompts from JailBreakV-28K. Every prompt is assigned to one of 14 cybersecurity attack categories (e.g., malware, phishing, privilege escalation) via a six-model majority-vote pipeline, and each strategy is ranked by effectiveness per category, enabling principled strategy selection grounded in concrete adversarial objectives. (2) Automated jailbreak generation. We instruction-fine-tune category-aware LLMs on Moderate and Optimal subsets, producing models that synthesize fluent jailbreak prompts from a harmful seed at inference time -- no templates, no gradient search. Our generators achieve perplexity 24-39 versus 40-140 for AutoDAN and AmpleGCG, with safety-filter evasion rates of 0.29-0.51 Mal (LlamaPromptGuard-2-86M), enabling controllable, scalable red-teaming under realistic adversarial conditions. (3) OPTIMUS: a training-free jailbreak evaluator. OPTIMUS is a continuous metric J(S,H) that jointly captures semantic similarity between the harmful seed and the jailbreak (S) and harmfulness probability (H) via calibrated penalty functions. Unlike binary attack success rate (ASR), OPTIMUS requires no task-specific training, generalizes across evolving strategies, and exposes a stealth-optimal regime (S*=0.57, H*=0.43) that ASR misses. Experiments across 114,000 prompts confirm that OPTIMUS separates Weak, Moderate, and Optimal jailbreaks with category-level evidence binary evaluation cannot supply.

Ismail Hossain, Tanzim Ahad, Md Jahangir Alam +3

Department of Computer Science, University of Texas at El Paso, TX, USA 79902 · School of Computing, Southern Illinois University Carbondale, IL, USA 62901 · University of Illinois Urbana-Champaign, IL, USA

Adversarial Prompt Generation Language Model Safety Evaluation LLM Jailbreak Attacks

May 9, 2026cs.CR

Single-Configuration Attack Success Rate Is Not Enough: Jailbreak Evaluations Should Report Distributional Attack Success

Many jailbreak attack research papers report attack success rates for a limited number of parameter settings, even though there are many combinations of parameter settings that could be used. Further, when new jailbreak papers are released, they often benchmark results against single configurations of existing attacks. This position paper argues such practices are fundamentally insufficient for characterising the threat posed by parameterised jailbreak attacks, and comparing attacks. Most jailbreak attacks expose multiple internal parameters, system prompt templates, conversation rounds, cipher dispersion, teaching shots, and ASR varies substantially across these parameters. Reporting only the best-case configuration discards two pieces of information that defenders genuinely need: how typical that performance is across the variant space, and how much of the attack surface is missed by selecting a single variant. We propose two new measures for jailbreak attacks: the Variant Sensitivity Measure (VSM) and Union Coverage (UC). VSM quantifies how far the best reported ASR deviates from the mean ASR across the tested variant space, UC is the total fraction of prompts resulting in unsafe responses across all tested configurations. We empirically demonstrate the importance of these measures using two attack families across three open-source target models. For PAIR, the best template reaches 69% ASR on Mistral-7B and 75% on Qwen3-0.6B, while UC rises to 88% and 93%, respectively. For bijection on Mistral-7B, the best variant reaches 81% ASR, but the 36-variant union covers 100% of HarmBench-100 prompts. We argue that distributional reporting, publishing VSM alongside ASR and enumerating variant coverage as fully as compute allows, should become the new minimum standard for parameterised jailbreak evaluation.

Carsten Maple, Abhishek Kumar, Riya Tapwal

University of Warwick · The Alan Turing Institute · IIT Mandi

Language Model Safety Evaluation Jailbreak Attacks

May 8, 2026cs.CR

Defense effectiveness across architectural layers: a mechanistic evaluation of persistent memory attacks on stateful LLM agents

Persistent memory in LLM agents creates an attack surface that production safety classifiers do not observe: the payload enters via RAG retrieval and persists across sessions via tool-mediated memory. We evaluate six defenses across four architectural layers against delayed-trigger attacks on nine open-source models (5,040 runs, N=40 per condition). Five of six defenses fail: input-level filters never see the payload (it enters via RAG, not user input); retrieval-level classifiers observe it but cannot distinguish compliance-framed injection from legitimate policy; instruction-level hardening is overridden by the stored rule's compliance framing. Only tool-gating at the memory layer (Memory Sandbox) reduces ASR to 0% for eight of nine models, with zero utility cost. A reasoning model inverts this defense via goal-directed RAG fallback, a mechanism that replicates cross-family on Bedrock. A reasoning-mode ablation reveals a double dissociation: no single sandbox implementation is safe across both reasoning and non-reasoning model classes. We resolve this with a content-layer proof-of-concept (RATG), validated on non-reasoning models. A loaded-corpus frontier evaluation (21 models, 3 providers, N=40) overturns an initial empty-corpus screen showing 0/210 exfiltrations: that was a threat-model artifact, not model safety. Under realistic conditions, Gemini 3.1 Pro Preview exfiltrates at 95% ASR, GPT-5.1 regresses to 22.5% relative to GPT-5 (5%), and Anthropic blocks at the injection layer (0-17.5% storage, 0% ASR). Nearly all OpenAI and Gemini models store the rule at 100% regardless of execution resistance, creating supply-chain risk in shared-memory deployments. Defense effectiveness is determined by architectural layer and reasoning capability, not classifier quality.

Jun Wen Leong

Language Model Safety Evaluation AI Agent Security Agent Memory Poisoning Prompt Injection Defense RAG Security Prompt Injection Attacks on LLMs

May 8, 2026cs.AI

The Attacker in the Mirror: Breaking Self-Consistency in Safety via Anchored Bipolicy Self-Play

Self-play red team is an established approach to improving AI safety in which different instances of the same model play attacker and defender roles in a zero-sum game, i.e., where the attacker tries to jailbreak the defender; if self-play converges to a Nash equilibrium, the model is guaranteed to respond safely within the settings of the game. Although the parameter sharing enforced by the use of the same model for the two roles improves stability and performance, it introduces fundamental theoretical and architectural limitations. We show that the set of Nash equilibria that can be reached corresponds to a broad class of behaviours that includes trivial always refuse strategies and oracle-like defenders, thus limiting practical applicability. We then show that when attacker and defender share and update the same base model, the dynamics collapse to self-consistency, so that attacks do not enforce adversarial pressure on the defender. In response, we propose Anchored Bipolicy Self-Play, which trains distinct role-specific LoRA adapters on top of a frozen base model, thereby maintaining stable optimisation while preserving adversarial pressure through explicit role separation. In relation to standard self-play, we show up to 100x greater parameter efficiency than finetuning and consistent improvements in safety compared to self-play fine-tuned models. We evaluate on Qwen2.5-{3B, 7B,14B}-IT models across widely used safety benchmarks, showing improved robustness without loss of reasoning ability. Cross-play experiments further show that our attacker and defender models are superior to self-play in terms of adversarial defence and safety.

Gabriele La Malfa, Emanuele La Malfa, Saar Cohen +4

Department of Informatics, King’s College London · Department of Computer Science, University of Oxford · Institute for Decentralized AI +1

Self-Play RL Language Model Safety Evaluation Adversarial Attacks on LLMs LLM Safety Alignment Adversarial Robustness of Language Models

May 7, 2026cs.LG

A Systematic Investigation of RL-Jailbreaking in LLMs

The evolution of generative models from next-token predictors to autonomous engines of complex systems necessitates rigorous safety hardening. Adversarial jailbreaking, the strategic manipulation of models to elicit harmful output, remains a primary threat to safe deployment. While Reinforcement Learning (RL) frames jailbreaking as a multi-step attack through sequential optimization, a mechanistic understanding of why the framework succeeds remains incomplete. To fill this gap, we present the first systematic decomposition of RL jailbreaking. We deconstruct the framework into problem formalization (reward function, action space, episode length), and algorithmic measures (RL algorithm, training data, reward-shaping) to identify the structural determinants of adversarial success. Our results reveal that the RL-jailbreaker successfully compromised all targeted models and safeguards. Through this first-of-its-kind analysis, we demonstrate that environment formalization, specifically dense rewards and extended episode lengths, is the primary driver of jailbreaking success. This work provides a tool for improving RL-jailbreaker efficiency and, ultimately, harden generative models resistant to RL-based attacks.

Montaser Mohammedalamen, Kevin Roice, Reginald McLean +1

Alberta Machine Intelligence Institute (Amii), Edmonton, Canada.

Language Model Safety Evaluation Adversarial Attacks on LLMs LLM Jailbreak Attacks

May 7, 2026cs.LG

When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels

Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component (

η^2 \approx 0.52

), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.

Sushant Gautam, Finn Schwall, Annika Willoch Olstad +6

1Simula Metropolitan Center for Digital Engineering · 2Oslo Metropolitan University · 4Simula Research Laboratory +2

LLM Safety Benchmarks LLM Evaluation LLM-as-a-Judge Language Model Safety Evaluation LLM Auditing

May 7, 2026cs.CL

Algospeak, Hiding in the Open: The Trade-off Between Legible Meaning and Detection Avoidance

As large language models (LLMs) increasingly mediate both content generation and moderation, linguistic evasion strategies known as Algospeak have intensified the coevolution between evaders and detectors. This research formalizes the underlying dynamics grounded in a joint action model: when Algospeak increases, detectability and understandability decrease. Further, the concept of Majority Understandable Modulation (MUM) is introduced and defined as the modulation level at which additional evasive alteration increases detector evasion but loses comprehension for the majority of recipients. To empirically probe this trade-off, we introduce a reproducible framework that can be used to create meaning-preserving, Algospeak-style variants, based on an existing taxonomy and with tunable modulation levels. Using COVID-19 disinformation as a first proof-by-example setting, we construct a reference dataset of 700 modulated items, drawn from twenty base sentences across five modulation levels and seven strategies. We then run two linked evaluations with seven different language models: one testing for interpretation through meaning recovery and one for disinformation detection through classification. Curve fitting over modulation levels yields an estimate of the Majority Understandable Modulation threshold and enables sensitivity analyses across strategies and models, see Figure 1. Results reveal the characteristic relationships between understandability and modulation. This study lays the groundwork for understanding the dynamics behind Algospeak and provides the framework, dataset, and experimental setups described.

Jan Fillies, Ronald E. Robertson, Jeffrey Hancock

Stanford University Stanford, California, USA

Language Model Safety Evaluation Adversarial Text Attacks

May 7, 2026cs.LG

How Many Iterations to Jailbreak? Dynamic Budget Allocation for Multi-Turn LLM Evaluation

Evaluating and predicting the performance of large language models (LLMs) in multi-turn conversational settings is critical yet computationally expensive; key events -- e.g., jailbreaks or successful task completion by an agent -- often emerge only after repeated interactions. These events might be rare, and under any feasible computational budget, remain unobserved. Recent conformal survival frameworks construct reliable lower predictive bounds (LPBs) on the number of iterations to trigger the event of interest, but rely on static budget allocation that is inefficient in multi-turn setups. To address this, we introduce \emph{Dynamic Allocation via PRojected Optimization} (DAPRO), the first theoretically valid dynamic budget allocation framework for bounding the time-to-event in multi-turn LLM interactions. We prove that DAPRO satisfies the budget constraint and provides distribution-free, finite-sample coverage guarantees without requiring the conditional independence between censoring and event times assumed by prior conformal survival approaches. A key theoretical contribution is a novel coverage bound that scales with the square root of the mean censoring weight rather than the worst-case weight, yielding provably tighter guarantees than prior work. Furthermore, DAPRO can be employed to obtain unbiased, low-variance estimates of population-level evaluation metrics, such as the jailbreak rate, under limited computing resources. Comprehensive experiments across agentic task success, adversarial jailbreaks, toxic content generation, and RAG hallucinations using LLMs such as Llama 3.1 and Qwen 2.5 demonstrate that DAPRO consistently achieves coverage closer to the nominal level with lower variance than static baselines, while satisfying the budget constraint.

Shai Feldman, Yaniv Romano

Department of Computer Science Technion, Israel · Departments of Electrical & Computer Engineering. and Computer Science Technion, Israel

LLM Evaluation Language Model Safety Evaluation Survival Analysis

May 7, 2026cs.AI

Beyond Fixed Benchmarks and Worst-Case Attacks: Dynamic Boundary Evaluation for Language Models

Evaluating large language models (LLMs) today rests on fixed benchmarks that apply the same set of items to any model, producing ceiling and floor effects that mask capability gaps. We argue that the most informative evaluation signal lies at the boundary, where the per-prompt pass probability is near

0.5

under random-sampling decoding strategy, and propose Dynamic Boundary Evaluation (DBE), which actively locates each model's boundary and places it on a globally comparable difficulty scale. DBE delivers three artifacts: (i) a calibrated item bank covering safety, capability, and truthfulness, with per-item difficulty labels validated across

9

reference LLMs; (ii) Skill-Guided Boundary Search (SGBS), a search algorithm that finds boundary items for a given target LLM using only API-level query access; and (iii) an evaluation protocol that places a new LLM on a unified ability scale and grows the evaluation set adaptively when the target falls outside the bank's coverage. We instantiate DBE on four categories spanning safety (harmful request refusal, over-refusal), capability (constrained instruction following), and truthfulness (multi-turn sycophancy resistance). The resulting evaluation covers a broader model spectrum without saturation while remaining compatible with existing datasets.

Haoxiang Wang, Da Yu, Huishuai Zhang

Peking University · Google Research · Wangxuan Institute of Computer Technology, Peking University State Key Laboratory of General Artificial Intelligence

LLM Evaluation Language Model Generation Evaluation Language Model Safety Evaluation

May 7, 2026cs.CL

Evaluation Awareness in Language Models Has Limited Effect on Behaviour

Large reasoning models (LRMs) sometimes note in their chain of thought (CoT) that they may be under evaluation. Researchers worry that this verbalised evaluation awareness (VEA) causes models to adapt their outputs strategically, optimising for perceived evaluation criteria, which, for instance, can make models appear safer than they actually are. However, whether VEA actually has this effect is largely unknown. We tested this across open-weight LRMs and benchmarks covering safety, alignment, moral reasoning, and political opinion. We tested this both on-policy, sampling multiple CoTs per item and comparing those that spontaneously contained VEA against those that did not, and off-policy, using model prefilling to inject evaluation-aware sentences where missing and remove them where present, with subsequent resampling. VEA has limited effect on model behaviour: injecting VEA into CoTs produces near-zero effects (

ω\leq 0.06

), removing it causes small shifts (

ω\leq 0.12

) and spontaneously occurring VEA shifts answer distributions by at most 3.7 percentage points (

ω\leq 0.31

). Our findings call for caution when interpreting high VEA rates as evidence of strategic behaviour or alignment tampering. Evaluation awareness may pose a smaller safety risk than the current literature assumes.

Amelie Knecht, Lucas Florin, Thilo Hagendorff

University of Stuttgart

LLM Alignment Language Model Safety Evaluation Evaluation Awareness in Language Models

May 7, 2026cs.CL

XL-SafetyBench: A Country-Grounded Cross-Cultural Benchmark for LLM Safety and Cultural Sensitivity

Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.

Dasol Choi, Eugenia Kim, Jaewon Noh +14

AIM Intelligence · Microsoft · Korea AISI +7

LLM Safety Benchmarks Multilingual Language Model Evaluation Language Model Safety Evaluation Cross-Cultural Language Model Evaluation Adversarial Attacks on LLMs Cultural Bias in Language Models+2

May 7, 2026cs.CL

One Turn Too Late: Learning When to Intervene Against Multi-Turn Malicious Intent

Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, attackers can distribute their intent across multiple benign-looking turns, making defense a problem not only of whether a dialogue is harmful, but also of when intervention becomes necessary. Existing trace-level labeling approaches provide only coarse safety signals and do not identify this intervention boundary, making it difficult to distinguish timely intervention from premature refusal or a block that comes too late. This work introduces turn-level harm-enabling supervision for multi-turn defense. We define the earliest harm-enabling turn as the first point at which delivering a candidate response would make the accumulated interaction sufficient to enable harmful action. To instantiate this supervision at scale, we construct the Multi-Turn Intent Dataset (MTID), which contains adaptive attack rollouts, matched benign hard negatives, and annotations of this boundary. Using MTID, we train TurnGate, a response-aware monitor that learns when to intervene, and further optimize its policy through multi-turn reinforcement learning. Experiments show that turn-level boundary supervision improves intervention localization, while reinforcement learning further improves the safety--utility trade-off. TurnGate outperforms existing guardrails and multi-turn monitoring baselines, and generalizes across risk domains, attacker pipelines, and target models. Our code is available at https://github.com/Graph-COM/TurnGate.

Xinjie Shen, Rongzhe Wei, Peizhi Niu +6

Georgia Institute of Technology · University of Illinois Urbana-Champaign · UCSD +3

Language Model Safety Evaluation LLM Guardrails Intent Classification LLM Safety Multi-Turn Jailbreak Attacks

May 6, 2026cs.AI

The Refusal--Compliance Tradeoff: A Large-Scale Safety Behavior Audit of Large Language Models

Refusal rates are a poor proxy for LLM safety, i.e., a model may over-refuse benign prompts while still complying with harmful ones. We audit both failure modes across 21 open-weight LLMs on four safety benchmarks (OR-Bench, XSTest, ToxiGen, BOLD), using a composition adjustment to isolate model sensitivity from dataset toxicity confounds. We report three findings. First, models adopt fundamentally different calibration strategies: conservative ecosystems such as Llama suppress unsafe outputs at the cost of elevated over-refusals, while permissive ecosystems such as DeepSeek and Qwen preserve helpfulness but tolerate higher harmful compliance. Second, demographic protection is unequal: models over-protect prominent racial and religious groups, frequently refusing even benign prompts about them, while providing substantially weaker protection against disability-targeted attacks. Third, refusal and compliance tendencies are stable within model families across generations and scales, suggesting that post-training objectives shape safety behavior more than architecture. Our results call for joint, demographically-aware, and multi-judge safety evaluation.

Alif Al Hasan, Sumon Biswas

Department of Computer and Data Sciences Case Western Reserve University Cleveland, OH, USA

LLM Safety Benchmarks Language Model Safety Evaluation LLM Refusal Behavior LLM Safety Over-Refusal Demographic Bias in Language Models+1

May 6, 2026cs.CR

SoK: Robustness in Large Language Models against Jailbreak Attacks

Large Language Models (LLMs) have achieved remarkable success but remain highly susceptible to jailbreak attacks, in which adversarial prompts coerce models into generating harmful, unethical, or policy-violating outputs. Such attacks pose real-world risks, eroding safety, trust, and regulatory compliance in high-stakes applications. Although a variety of attack and defense methods have been proposed, existing evaluation practices are inadequate, often relying on narrow metrics like attack success rate that fail to capture the multidimensional nature of LLM security. In this paper, we present a systematic taxonomy of jailbreak attacks and defenses and introduce Security Cube, a unified, multi-dimensional framework for comprehensive evaluation of these techniques. We provide detailed comparison tables of existing attacks and defenses, highlighting key insights and open challenges across the literature. Leveraging Security Cube, we conduct benchmark studies on 13 representative attacks and 5 defenses, establishing a clear view of the current landscape encompassing jailbreak attacks, defenses, automated judges, and LLM vulnerabilities. Based on these evaluations, we distill critical findings, identify unresolved problems, and outline promising research directions for enhancing LLM robustness against jailbreak attacks. Our analysis aims to pave the way towards more robust, interpretable, and trustworthy LLM systems. Our code is available at Code.

Feiyue Xu, Hongsheng Hu, Chaoxiang He +9

Shanghai Jiao Tong University, China · Microsoft Corporation

Language Model Safety Evaluation LLM Security LLM Jailbreak Attacks Adversarial Robustness of Language Models Jailbreak Defense

May 6, 2026cs.CL

Misaligned by Reward: Socially Undesirable Preferences in LLMs

Reward models are a key component of large language model alignment, serving as proxies for human preferences during training. However, existing evaluations focus primarily on broad instruction-following benchmarks, providing limited insight into whether these models capture socially desirable preferences. As a result, important failures in social alignment can remain hidden. We extend reward-model benchmarking to four socially consequential domains: bias, safety, morality, and ethical reasoning. We introduce a framework that converts social evaluation datasets into pairwise preference data, leveraging gold labels where available and directional bias indicators otherwise. This enables us to test whether reward models prefer socially undesirable responses, and whether their preferences produce systematically biased distributions over selected outputs. Across five publicly available reward models and two instruction-tuned models used as reward proxies, we find substantial variation across domains, with no single model performing best overall. The models fall well short of strong social intelligence: they often prefer socially undesirable options, and their preferences produce systematically biased distributions. Moreover, stronger bias avoidance can reduce sensitivity to context, revealing a key alignment trade-off between avoiding biased outcomes and preserving contextual faithfulness. These findings show that standard reward benchmarks are insufficient for assessing social alignment and highlight the need for evaluations that directly measure the social preferences encoded in reward models.

Gayane Ghazaryan, Esra Dönmez

Institute for Natural Language Processing, University of Stuttgart · 2Interchange Forum for Reflecting on Intelligent Systems, University of Stuttgart

Reward Modeling Language Model Safety Evaluation Social Bias in Language Models Preference Alignment Reward Model Evaluation

May 5, 2026cs.CR

MOSAIC-Bench: Measuring Compositional Vulnerability Induction in Coding Agents

Coding agents often pass per-prompt safety review yet ship exploitable code when their tasks are decomposed into routine engineering tickets. The challenge is structural: existing safety alignment evaluates overt requests in isolation, leaving models blind to malicious end-states that emerge from sequenced compliance with innocuous-looking requests. We introduce MOSAIC-Bench (Malicious Objectives Sequenced As Innocuous Compliance), a benchmark of 199 three-stage attack chains paired with deterministic exploit oracles on deployed software substrates (10 web-application substrates, 31 CWE classes, 5 programming languages) that treats both exploit ground truth and downstream reviewer protocol as first-class evaluation axes. On this benchmark, nine production coding agents from Anthropic, OpenAI, Google, Moonshot, Zhipu, and Minimax compose innocuous tickets at 53-86% end-to-end ASR with only two refusals across all staged runs. In a matched direct-prompt experiment over four frontier Claude/Codex agents, vulnerable-output rates fall to 0-20.4%: Claude primarily refuses, while Codex primarily hardens rather than emitting the vulnerable implementation - ticket staging silences both defense modes simultaneously. Downstream, code reviewer agents approve 25.8% of these confirmed-vulnerable cumulative diffs as routine PRs, and a full-context implementation protocol closes only 50% of the staged/direct gap, ruling out context fragmentation as the sole explanation. As a deployable but non-adaptive mitigation, reframing the reviewer as an adversarial pentester reduces evasion across the evaluated reviewer subset; pentester framed evasion ranges from 3.0% to 17.6%, and an open-weight Gemma-4-E4B-it reviewer under this framing detects 88.4% of attacks on the dataset with a 4.6% false-positive rate measured on 608 real-world GitHub PRs.

Jonathan Steinberg, Oren Gal

Swarms & AI Lab (SAIL), University of Haifa

Language Model Safety Evaluation AI Coding Agents Adversarial Attacks AI Agent Security Benchmarks Automated Code Review AI Agent Safety Benchmarks