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

Oct 8, 2026cs.AI

Searching for "Harmful Refusal": A Psychometric Audit of an AI Safety Benchmark

Safety benchmarks typically report one overall score for a suite of datasets, each of which may target one or more safety-related attributes, so models with similar overall scores can have very different attribute profiles. Comparing models is more tractable at the level of individual attributes, yet it is often unclear whether even a single dataset's scores isolate any single attribute. One plausible candidate for such an attribute is harmful refusal, a model's tendency to refuse dangerous or policy-violating prompts. We examine whether it constitutes a single, measurable attribute in HELM Safety. Using a construct validity framework that stipulates that an attribute must exist before a test can measure it, we start with HELM Safety's four datasets that might plausibly target harmful refusal, but find that three are saturated. We subject the remaining dataset, HarmBench, to two psychometric tests to determine if a single attribute like harmful refusal could stand behind its score. First, multidimensional item response theory modeling strongly suggests that HarmBench does not measure a singular attribute. Second, a differential item functioning analysis finds items where models from different developers with the same refusal ability score differently. These flags largely disappear under scope-specific matching, a pattern consistent with aggregation effects but not sufficient to rule out domain-specific developer differences. Zooming out, HarmBench collapses distinct harm behaviors into one score, and the overall HELM safety aggregate further collapses HarmBench and scores from other datasets into a single top-line number. Any safety score that averages over datasets and items can hide saturation and conflate behaviors this way. We argue that a score should earn its single-attribute reading before models are compared with it.

Christopher M. Stewart, Preston Botter, Natalie Sarabosing +5

Carnegie Mellon University · Indiana University · Google

LLM Safety Benchmarks Language Model Safety Evaluation LLM Evaluation LLM Refusal Behavior

Oct 8, 2026cs.AI

One Word Opens the Gate: The Option-Channel Attack on Typed Decision Models as Agent Guardrails

A typed decision model reads a piece of text and returns a probability over caller-defined options, each with a short written definition, generating no text. Recent work places these models in agent systems as guardrails: the component that reads a proposed tool call or incoming message and decides whether to allow it. We evaluate seven open-weight models in that role and report the two error directions separately: a fail-open error allows a prohibited action and is a vulnerability; a fail-closed error blocks a permitted one and is only a cost. On prompt-injection, jailbreak and toxic-content screening, accuracy at the allow-or-block decision ranges from 36% to 72% against a chance level of 50%. A low error rate in one direction only reflects which answer a model defaults to: one allows nearly everything, another blocks nearly everything. On a synthetic suite of agent tool calls, six lines of server log text that say nothing about the policy raise a gate's fail-open rate from 0% to 63% on a policy it otherwise decides correctly. Giving the permissive option a misleading name, with its definition and the judged text untouched, raises that rate to between 93% and 100% on the four models that place the label in their input. Every defense we tested is defeated, either by an attacker who targets its mechanism or by attacker-controlled text. Escalating the least confident decisions does not help either: a decision an attack has reversed is no less confident than the one it replaced. Parsing each policy field into a typed value does eliminate one attack, but it also makes the model unnecessary: a deterministic rule over those values reaches 100% accuracy on all six policies. These models can reduce how many cases reach a reviewer, but on this evidence they should not be the component that decides. Code is available at https://github.com/ArminAzizi98/option-channel-attack.

Seyedarmin Azizi, Erfan Baghaei Potraghloo, Massoud Pedram

University of Southern California (USC), USA

LLM Guardrails LLM Agent Security Adversarial Attacks on LLMs Language Model Safety Evaluation Prompt Injection Attacks on AI Agents

Oct 8, 2026cs.CR

ReSI: Recursive Safety Improvement toward Resistant and Resilient AI

Recursive self-improvement, the participation of AI systems in improving their own capabilities, is beginning to move from theoretical prospect to practice, posing both challenges and opportunities for safety alignment. Models evolve through frequent updates, and their safety alignment requires continual adaptation to each new checkpoint. Meanwhile, with evolving red-teaming methods exposing new vulnerabilities, safety improvement for each checkpoint needs to mitigate exposed vulnerabilities and generalize to risks not yet revealed. Following R

^2

AI, we term these goals resistance to known threats and resilience to unforeseen risks. Recursive self-improvement, in turn, inspires an approach to both goals: safety alignment could likewise advance through successive rounds of evaluation and update. We therefore introduce ReSI, a recursive safety improvement framework that implements this approach through automated research. In each round, ReSI applies diverse red-teaming methods to identify vulnerabilities in the current target model, develops training recipes, and promotes the update with the largest safety gain among those passing a Pareto gate on capability retention as the next target model. Across four dense and mixture-of-experts models, ReSI matches or exceeds evaluated frontier models on in-distribution and out-of-distribution safety benchmarks, and outperforms alignment baselines on nearly all safety evaluations while largely preserving general capabilities. In particular, ReSI reduces the mean X-Teaming attack success rate across the four models from 86.01% to 31.45%, well below GPT-5.6-Luna's leading frontier result of 56.69%, indicating stronger resilience to attacks unseen during training. These findings support recursive safety improvement as a practical path toward resistant and resilient AI.

Jingnan Zheng, Dongcheng Zhang, Yi Zhang +10

Shanghai AI Laboratory · National University of Singapore · Shanghai Jiao Tong University +2

LLM Safety Alignment Recursive Self-Improvement Language Model Safety Evaluation

Oct 8, 2026cs.CL

Same Outcome, Different Evidence: Intent Recovery in LLM Safety Evaluation

Safety evaluations of large language models commonly summarize harmful-output behavior with attack success rate (ASR). Yet the same non-harmful outcome can arise for very different reasons. A model may recover a harmful task and refuse it, fail to recover the task, or respond to something else entirely. Distinguishing these cases becomes especially important under intent-obscuring prompts, where a low ASR does not reveal whether the evaluated task was actually engaged. To make this distinction explicit, we pair ASR with operative understanding rate (UR), which measures whether a response both identifies the evaluated task and treats it as the task to be answered. Across interfaces, this paired view reveals substantial variation hidden by ASR: similar ASR values can correspond to sharply different recovery rates. Controlled English reconstructions show that recovery consistently improves as compressed prompts become more explicit, whereas ASR does not follow the same pattern. A complementary contrast comes from FormalLogic, where high recovery can still coincide with frequent harmful assistance. Together, these results show that non-harmful outcomes are not equally informative about model safety, motivating the joint reporting of intent recovery and ASR in LLM safety evaluation. Code and experiment inputs are available at https://github.com/kevinjiang0121-cyber/IRIS.

Haitong Jiang, Chunlin Liu, Sihan Tang +3

College of Computer Science and Software Engineering, Shenzhen University · School of Intelligence Science and Engineering, Harbin Institute of Technology (Shenzhen), Shenzhen, China

Language Model Safety Evaluation

Oct 7, 2026cs.AI

How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense

Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper. We measure this vulnerability across languages and registers: attack success on Qwen3-1.7B is already 89.4% in English and 93.0% in modern Chinese, and reaches 95.7% in Classical Chinese. We build GUISE, a benchmark for systematically studying this vulnerability. It includes parallel requests in English, modern Chinese, and Classical Chinese, matched harmful and benign pairs, wrapper types held out for evaluation, and a stricter criterion that counts warn-then-answer responses as attack successes. Representation analysis shows that language and register move harmful-request representations only slightly away from the model's refusal direction, whereas narrative wrappers move them much farther away. We propose AXIS, which combines preference optimisation with a rotation objective that aligns harmful-request representations with the refusal direction and a commitment objective that trains the model to refuse completely rather than produce a warn-then-answer response. Across Qwen3-1.7B, Qwen3-4B and GLM-4-9B, AXIS achieves the highest combined safety and usability score among the compared methods.

Zhankai Ye, Yanning Wang, Yukai Jin +5

Florida State University · Texas Christian University · University of Pennsylvania +1

Refusal Behavior in Language Models LLM Safety Alignment LLM Jailbreak Attacks Language Model Safety Evaluation LLM Safety Benchmarks

Oct 7, 2026cs.LG

PatchBench: Measuring Collateral Damage in Activation Patching

An LLM safety patch can pass a benchmark while still being a poor repair. This risk is especially acute for jailbreak repairs, where the goal is to correct a specific unsafe behaviour without changing unrelated behaviours. A patch may block exact evaluation prompts yet fail on close harmful variants, or suppress harmful behaviour by over-refusing benign prompts that share its wording or structure. Existing protocols primarily test whether models can be broken, while aggregate metrics (attack success, refusal rates, global capability) cannot distinguish selective repairs from broader local suppression. To address this gap, we introduce PatchBench, a benchmark of empirically observed model-specific jailbreak failures inducing actionable harmful answers. Starting from 27,870 prompts from 37 public datasets, we curate 15,314 English prompts and query 8 open-source instruction-tuned models. Combining WildGuard filtering, pairwise Elo ranking, and manual verification, we retain a curated bank of 400 high-confidence jailbreak failures. We further introduce PatchBench-Local, an evaluation protocol testing whether a patch is behaviourally precise. For each harmful source prompt, PatchBench-Local generates three families of local neighbours: harmful variants preserving malicious intent, benign prompts with matched structure, and benign prompts reusing key harmful terms. It evaluates harmful-neighbour correction and benign-neighbour preservation, distinguishing selective repair from broader local suppression. Evaluating four activation steering methods with PatchBench-Local and MMLU shows that global capability can remain nearly unchanged while local benign regressions are severe, confirming aggregate metrics miss important collateral damage. PatchBench-Local provides a more precise basis for developing and comparing jailbreak repair methods.

Alexi Canesse, Mathis Le Bail, Maël Jenny +3

LIX (École Polytechnique, IP Paris, CNRS), France · AMIAD (Agence Ministérielle pour l’IA de Défense), France

Language Model Safety Evaluation Jailbreak Defense Jailbreak Attacks LLM Jailbreak Attacks Activation Steering

Oct 6, 2026cs.CL

Quad-State Safety Evaluation of Open-Weight Large Language Models on Non-Canonical Inputs

Standard safety evaluations of large language models assess harmful requests written in canonical plain text, while models in real-world deployment routinely receive inputs containing emojis, altered spellings, encoded strings, and character-level variations. This work introduces the Adversarial Surface-Form Robustness Dataset (ASRD), comprising 2,100 prompts across seven distinct surface-form families. Five open-weight language models are evaluated across these prompts, producing 10,500 responses. The Quad-State Evaluation Rubric classifies each response into one of four outcomes: harmful compliance, safe response, comprehension failure, or indeterminate. Emoji and invisible Unicode variations cause almost no comprehension failure, with pooled harmful compliance of 20.27% and 17.20% against a 22.87% baseline that is driven mainly by Mistral 7B, whereas leetspeak, encoded wrappers, and hybrid transformations score 2.40%, 0.13%, and 2.40% while comprehension failure rises to 36.47%, 65.60%, and 34.47%. Inspection of raw model outputs reveals three response behaviors: hallucinated benignity, structural collapse, and language drift. Project page: www.pavanmaddula.com/quadstate

Pavan Maddula

Independent Researcher

Language Model Safety Evaluation Adversarial Robustness of Language Models LLM Safety Benchmarks Adversarial Attacks on LLMs

Oct 6, 2026cs.AI

Not Every Call Needs a Frontier Model: Per-Call-Site Evaluation of Small Language Models in a Deployed Agentic Home-Automation System

An agentic system issues several structurally different kinds of LLM calls. It routes intent, classifies actions, grounds language in a device registry, plans multi-agent pipelines and writes the Python code those pipelines run. The difficulty of these call sites varies by an order of magnitude, yet in practice a single model, chosen for the hardest site, serves all of them. In this work, we evaluate 9 models from 0.8B to a frontier hosted model across the five call sites of a deployed open-source home-automation framework (Wactorz), using its unmodified production prompts and two real Home Assistant installations (280 cases, 2520 scored calls). We find that capability is not ordered the same way at every site, and that larger models are not uniformly better: one 4B model is worse than its 2B sibling at grounded actuation. Paired testing shows the best local model to be statistically indistinguishable from both hosted models at four of five sites. Only code generation separates them, against a small hosted model (p = 0.039) as well as a frontier one (p = 0.002). Aggregate accuracy also hides a safety failure specific to actuation, where small models resolve the accuracy/refusal trade-off in degenerate ways: one model (Gemma4 E2B) actuates on 87.2% of requests for devices the site does not own, while another refuses every request it receives. Routing each site to its best local model reaches 91.8% against 95.4% at no per-call cost. In a live deployment judged by a user, hosting only the two generative sites matches hosting everything (39/43 against 39/43) for 28% of the spend, and the actuation gap the benchmark predicted appears as exactly one case in twenty-six. Benchmark, harness and all records are released at https://github.com/waldiez/slm-callsite-eval.

Panagiotis Kasnesis, Christos Chatzigeorgiou, Lazaros Toumanidis +1

University of West Attica, Athens, Greece · Waldiez PC · ThinGenious PC

Small Language Models LLM Agent Evaluation LLM Routing Language Model Safety Evaluation

Oct 6, 2026cs.AI

How Fragile Is On-Device Language Model Safety? Localizing Safety-Critical Parameters for Sparse Fault Analysis

As small language models (SLMs) are increasingly deployed on resource-constrained and on-device platforms, including as components of agentic systems, the integrity of locally stored model parameters becomes an important safety concern. We investigate whether safety-sensitive behavior in LLaMA-2-7B-Chat is concentrated within a sparse subset of parameters, creating a reduced fault surface for targeted analysis. We study two complementary localization methods: low-rank safety-associated subspace analysis and parameter-level safety--utility importance filtering. Both approaches reveal highly non-uniform safety sensitivity across the network, with the MLP down_proj consistently emerging as a prominent safety-sensitive component and o_proj providing a smaller contribution. Using parameter-level localization, modifying only 0.19% of model weights in down_proj yields 53% Basic ASR and 56% GCG ASR, while tinyBenchmarks accuracy remains at 51.6% compared with a 52.2% unmodified baseline. These results motivate targeted fault analysis and selective integrity protection for language models deployed in resource-constrained, on-device, and agentic settings.

Muhammad Zeeshan Karamat, Christiana Chamon Garcia

Virginia Tech

LLM Safety Language Model Safety Evaluation Adversarial Attacks on LLMs

Oct 6, 2026cs.AI

Toward Alignment Scaling Laws: A Framework and First Preregistered Measurements

Whether alignment gets easier or harder as models grow is often argued from isolated findings, as if alignment were one property. We treat it as a family of measurable scaling relations: for each risk category r, the alignment burden needed to hold a fixed safety target is modeled as B_r(N)=a_rN^alpha_r, with N a capability proxy; against a budget proportional to N, scaling helps if alpha_r<1, keeps pace if alpha_r~1, and accumulates alignment debt if alpha_r>1. We give three operationalizations of burden and distinguish observed, audited and true alignment. A toy model, in which corrections consume capability headroom, makes the consequences explicit. We prove that the largest exponent among corrected risks, not an average, sets the long-run regime; that above 1 any policy holding headroom above a floor must grow super-exponentially; that, for burdens that are positive mixtures of power laws, fits on small models underestimate large-scale exponents; and that an audit that uncovers hidden failures without false positives never underestimates true alignment. We propose a pre-registrable protocol and apply reduced versions of it twice. A preregistered reanalysis of public adversarial-training data for Pythia classifiers finds that the compute needed to bring attack success under 10% grows as N^0.60. A preregistered pilot on Qwen2.5 0.5B-72B finds exponents of -0.05 for truthfulness and 0.48 for stated dispositions (both scaling helps under its reduced rule, though local slopes approach 1 at the top; replicated on Qwen3 0.6B-14B), while sycophancy (0.89, or 0.83 with two seeds added at 72B) and a planted backdoor are undetermined: the backdoor is removed quickly when its trigger is known but survives blind safety training at four of five sizes. We release four browser games that play these laws (www.aisafety.fun). We make no claim about which regime holds for current frontier models.

Jeremy Canale

LLM Alignment Language Model Safety Evaluation Language Model Scaling Laws

Oct 6, 2026cs.CL

Does Steering Break Your Model? A Multi-Dimensional Evaluation Suite for LLM Steering Methods

Activation steering provides a lightweight and flexible way to control large language model (LLM) behavior. However, effective steering requires more than inducing the intended behavior: it should also limit unintended changes and remain robust across inputs and training data. Existing evaluations cover these dimensions only in fragments. As a result, the trade-offs between efficacy and side effects have not been systematically characterized. We introduce SteerScope, a two-axis, multi-dimensional evaluation suite that jointly characterizes steering outcomes and method properties through 15 metrics. We score target efficacy and side effects on language quality, task capabilities, and safety and reliability, and further assess generalization and data dependence through steering-specific metrics for sample efficiency and sample sensitivity. Rather than comparing methods at a single operating point, we characterize the trade-offs between efficacy and side effects. Under matched models, tasks, and evaluation protocols, we benchmark 23 methods spanning 4 families, including prompting, LoRA, and SFT as baseline methods, and release the suite as an extensible codebase. We find that current activation steering methods do not yet surpass the Prompt Steering baseline in their overall balance between steering efficacy and side effects: across both model scales, no evaluated activation steering method achieves higher efficacy without incurring greater composite side effects. We further uncover a consistent coupling between steering efficacy and side effects. Under OOD prompts, target efficacy is often preserved, whereas side effects tend to become more pronounced, particularly through declines in instruction relevance and fluency. Methods also exhibit sharply different sample-efficiency profiles.

Haotian Yang, Huikang Jiang, Yucheng Wu +4

State Key Laboratory of Multimedia Information Processing, Peking University · School of Computer Science, Peking University · Columbia University +2

LLM Evaluation Language Model Steering Language Model Safety Evaluation Activation Steering

Oct 5, 2026cs.AI

The Pushback Paradox: A Two-Probe Diagnostic for Language Model Compliance

Are language models compliant with user instructions? A model that always complies can be stopped but also exploited, while one that always resists can be neither exploited nor stopped. We contribute an open two-probe benchmark that can place any language model on this spectrum. In the active probe, a user instructs the model to act and accept a lower payoff, which measures exploitability. In the passive probe, the user instructs it to wait and give up a higher payoff, which measures stoppability. The two compliance rates combine into a compliance index

κ

. Applied to twelve language models, the benchmark shows that seven mostly follow the instruction in both probes and justify their action by pointing to the instruction. Only Claude Sonnet-4.6 and Claude Opus-4.7 can be stopped without being exploitable, Claude Opus-4.6 and GPT-5-mini resist both instructions, and no model is exploitable but unstoppable. Knowing where a language model sits on the compliance index

κ

matters for human operators and for multi-agent systems, whether distributed or orchestrated.

Stefan Bühler, David Exler, Markus Reischl +1

Independent Researcher · Institute for Automation and Applied Informatics, Karlsruhe Institute of Technology, Karlsruhe · Duale Hochschule Baden-Württemberg, Ravensburg

LLM Safety Benchmarks Language Model Safety Evaluation

Oct 5, 2026cs.RO

What the Guard Misses, the Robot Executes: Implied Harm in VLA Instructions

Vision-language-action models (VLAs) act on instructions without being able to refuse, so screening harmful requests falls to monitors. We test whether these monitors catch ordinary robot tasks requested for harmful reasons, holding the task fixed while varying only how explicitly the intent is stated.

π_{0.5}

completes the task at every level of explicitness, as often as for harmless controls. Text guards flag nearly every blunt request but few implied ones: up to 95% of implied-harm runs end with the task done and no flag raised, and up to 90% even after recalibrating on robot instructions. Monitoring the model's activations does not close this gap. Linear probes separate harmful from harmless instructions almost perfectly in the base language and vision-language models, but this separation weakens after robot training in two model families, most for implied harm.

Sripad Karne, Arjun Balaji

Columbia University

Language Model Safety Evaluation Robot Safety Vision-Language-Action Models Safety Evaluation of VLA Models

Oct 4, 2026cs.AI

AI Safety via Debate is Compromised by Cognitive Biases

Reinforcement learning from human feedback (RLHF) has played a central role in making large language models responsive to human instructions. However, human evaluators often favor flattering or persuasive responses over truthful ones, creating incentives for models to appeal to evaluators at the expense of accuracy. AI safety via debate has been proposed as a way to improve the supervision of language models: in this paradigm, two agents argue opposing positions and challenge each other's claims, potentially exposing falsehoods to the adjudicator. A central premise of AI safety via debate is that truthful arguments are easier to defend than false ones under adversarial scrutiny. In this work, we investigate whether this advantage persists when debaters use rhetorical strategies that exploit biases in human judgment. Inspired by competitive debate, we construct 68 LLM-generated dialogues about detective mysteries with known culprits, spanning four interventions: anchoring, fallacy oversight, pro-jargon, and verbosity. We apply each intervention to either the side advocating for the true culprit or the side advocating for an innocent suspect, allowing us to distinguish influence on adjudication from correctness. In a study with 369 participants, we find that, pooled across bias types, these interventions significantly shift judgments toward the manipulated side. These findings expose a vulnerability in debate-based supervision: human adjudication is sensitive to manipulative rhetorical strategies.

Gefei Liu, Sonya Rashkovan, Sophia Lloyd George +2

Brown University

Language Model Safety Evaluation AI Safety

Oct 1, 2026cs.LG

A Safe Prototype Is Not a Safety Direction: Reference Dependence and Prompt Confounds in Response-Safety Embeddings

Can response safety be scored by cosine similarity to the mean embedding of known-safe responses? A recent sleeper-agent detector proposes exactly this score, yet the raw positive-centroid rule is not identified: positive observations locate the safe class relative to an encoder origin, but do not determine which direction separates safe from unsafe responses. We audit the rule on two prompt-controlled, human-labeled corpora and one auxiliary jury-labeled source control, using four frozen encoders and prompt-grouped splits. On the human-labeled corpora the safe prototype reaches ROC-AUC 0.457-0.545, with two cells significantly below chance and one above, while an explicit safe-minus-unsafe reference reaches 0.588-0.738 on the same embeddings; on the jury control the prototype is inverted (0.358-0.405) and the reference reaches 0.754-0.793. At validation-calibrated 5% false-safe thresholds, the reference accepts more safe responses on PKU-SafeRLHF (0.153-0.263 versus 0.039-0.061 across encoders) and Aegis (0.189-0.291 versus 0.004-0.045), but not reliably on BeaverTails. A fully unlabeled held-out reference recovers part to most of the referenced ranking, much less when only 5% of the pool is unsafe, whereas 80-634 labeled unsafe responses recover most of it. Prompt-only ablations show that prompt-label composition can inflate uncontrolled evaluations. This is a bounded result about a raw positive centroid, not all one-class methods or safety-specialized guards. A class mean is a location, not necessarily a safety direction; a declared reference with enough unsafe mass identifies orientation.

Sahil Kadadekar

New York University

Prototype-Based Classification LLM Evaluation Language Model Safety Evaluation Text Embedding Evaluation Safety Filtering

Oct 1, 2026cs.CR

False Floors: LLM Safety Routing Evaluations Break Under Distribution Shift

Safety routers send each request to one of several models and are judged against the best single model. A major routing benchmark picks that comparator on the evaluation data. In the benchmark's own setting this is harmless, but under distribution shift it is not. On HELM Safety the selection cost is 0.003-0.030 of harm under random splits and 0.045-0.113 under held-out categories, comparable to the whole deficit attributed to routing, with its direction holding under either published judge alone. It rises seven- to ninefold on AgentDojo when suites are held out. Across seven safety corpora chosen by rules fixed in advance, three meet a registered interval test and four beat a later permutation null, and three of the four interval misses are corpora where some models have zero observed harm. Prior work proves the direction of this bias. We size it on harm and accuracy, show that it is larger under the held-out splits we measure, and bound it by optimism plus a shift-dependent regret. Scored honestly under shift, routing buys little on these benchmarks. In most pool cells the nested router serves the honest baseline's model, and on the nearly saturated AgentDojo corpus a perfect pre-dispatch router is worth at most two points of harm. We also find a model's expressed recognition of a late injection steerable. On held-out reruns an attacker who knows which model it faces lowers GPT-5.4's judged recognition by 19.6 points, confirmed by an independent label. In an offline counterfactual composition into a controller, the same attack raises or lowers estimated harm depending on the fallback model. Safety routing should be evaluated under shift, against a baseline chosen without the test labels, and recognition-based defences should be scored on harm against an attacker who chooses what the model sees.

Amit Singh Bhatti, Vishal Vaddina

Quantiphi Analytics

LLM Safety Benchmarks Distribution Shift Robustness Adaptive Model Routing Language Model Safety Evaluation Selection Bias

Sep 29, 2026cs.CL

ContextAdapt: Evaluating Contextual Adaptation and Value Alignment in LLMs

Values such as honesty, autonomy, and confidentiality are often regarded as general principles underpinning AI alignment. However, what it means to act in accordance with these values can depend on the context in which a decision is made. In this paper, we ask whether large language models (LLMs) appropriately adapt the application of a value across professional settings, while remaining consistent when contextual changes do not alter the relevant professional norm. To study this, we introduce ContextAdapt, an evaluation framework covering honesty, autonomy, and confidentiality across medicine, law, finance, and national security. Drawing on primary-source professional and regulatory documents, we construct a value x domain framework and use this to develop scenarios testing both default professional rules and recognised exceptions. We evaluate 12 LLMs on both the actions they recommend and the justifications they provide. In our main experiment, models achieve 95.6% mean appropriateness, although the use of the correct domain-specific justification varies substantially across models, from 25.6% to 76.9%. In a separate factorial experiment, explicitly naming the professional domain and changing the role of the model have limited effect on behaviour. Varying stakes, however, reveals severe but localised failures: in some cases, models alter their responses even though the underlying professional obligation remains unchanged. In particular, perceived severity appears to act as a cue for disclosure across both honesty and confidentiality scenarios. These results show that evaluating value alignment requires us to consider not only whether models follow abstract principles, but whether they apply them appropriately across different contexts.

Olivia Macmillan-Scott, Mirco Musolesi

Centre for AI, Department of Computer Science, University College London · Department of Computer Science and Engineering, University of Bologna

LLM Alignment Language Model Safety Evaluation

Sep 29, 2026cs.CR

Does the Unsafe Gradient Survive a Conversation? On the Fragility of Gradient-Based Jailbreak Detection in Multi-Turn Dialogue

Safety-aligned language models are commonly deployed as multi-turn assistants, which lets adversaries spread unsafe intent across several user turns instead of a single prompt. Gradient-based jailbreak detectors such as GradSafe were developed for single prompts: they score an input by the alignment between its induced gradient and a fixed unsafe reference direction, and their effectiveness in multi-turn dialogue remains unclear. We conduct a controlled evaluation of gradient-based jailbreak detection in multi-turn settings. We extend GradSafe with a Context Window Scanner that applies the detector to fixed-size windows of user turns and uses the maximum window score as the conversation-level score. We evaluate different window sizes, attack families, benign conversation distributions, and target models. The results differ sharply between synthetic and realistic benign settings. Against synthetic benign conversations, the detector achieves an ROC-AUC of 0.98 on human-authored multi-turn jailbreaks. On WildChat benign conversations, ROC-AUC drops to 0.76, and a threshold calibrated on synthetic data flags more than 90% of benign conversations as unsafe. Under realistic benign distributions, single-turn windows give the highest separability, whereas longer windows and accumulated contexts reduce performance. The detector is also sensitive to the attack-generation method and target model: successful Crescendo attacks receive scores comparable to or lower than benign conversations, and Qwen2.5-7B-Instruct yields near-random separability with a different optimal window size. These findings show that gradient-based signals can support multi-turn jailbreak detection, but reliable deployment requires calibration on realistic benign conversations, short-window scoring, length-aware thresholds, and evaluation across attack types and model architectures.

Omar Sheta, Rinku Deuja, Hadi Masoudi +1

University of Louisville Louisville, Kentucky, USA

Language Model Safety Evaluation Adversarial Attacks on LLMs Multi-Turn Dialogue Evaluation Jailbreak Detection Jailbreak Defense

Sep 29, 2026cs.LG

Guard Models Are Overconfident Where Base Models Are Uncertain

Guard models are used as safety classifiers, with confidence scores driving downstream moderation decisions. We evaluate five guard models for prompt classification and find that although several are nearly calibrated on clean inputs, adversarial attacks degrade their calibration by an order of magnitude, turning false negatives into high-confidence errors indistinguishable from correct detections. Comparing each guard with its corresponding base LM, we find that uncertainty signals often remain available, with the base model typically expressing uncertainty on the same inputs where the guard fails. Layer-wise analyses localize this guard-base divergence to later layers, where guard models exhibit sharper safe/unsafe separation and lower-rank representations, while adversarial harmful inputs lie closer to the clean-safe region. These findings highlight a mismatch between guard confidence and base model uncertainty under attack.

Jonghyun Hong, MinJae Jung, Minwoo Kim

DATUMO INC.

Language Model Safety Evaluation Adversarial Attacks on LLMs Language Model Calibration Safety Filtering LLM Safety LLM Uncertainty Estimation

Sep 28, 2026cs.AI

Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models

Large Reasoning Models (LRMs) are commonly trained with reinforcement learning (RL) to improve their generation of chain-of-thought (CoT) reasoning before producing final answers. However, RL rewards are typically assigned based on final answers, providing little or no direct supervision over intermediate reasoning. This can lead to deceptive safety alignment, where the reasoning trace and final answer convey inconsistent safety signals. To systematically investigate this phenomenon, we introduce DSAR (Deceptive Safety Alignment Rate), a metric that jointly assesses reasoning traces and final answers to quantify their safety inconsistency. Across multiple LRMs and benchmarks, we find that deceptive safety alignment is pervasive under standard prompting conditions and is substantially amplified under prefilling attacks. We further provide a hidden representation analysis showing that models exhibit stronger safety discrimination at the final-answer stage than during intermediate reasoning. To close this gap, we propose SARA (Safety-Aware Reasoning Alignment), an RL-based method that rewards both safety-aware reasoning and safe final answers, encouraging early harmful intent recognition and enforcing reasoning-answer consistency. Experiments show that SARA significantly mitigates deceptive safety alignment under both standard and adversarial settings while preserving helpfulness and utility. Code is available at https://github.com/xzhou98/SARA.

Xiangyu Zhou, Saleh Zare Zade, Rafi Ibn Sultan +2

Wayne State University

Language Model Safety Evaluation Deception in Language Models RL for Language Model Reasoning LLM Safety Alignment Large Reasoning Models

Sep 28, 2026cs.CL

Language Models Are "Insecure" Reporters

As large language models are deployed in increasingly autonomous long-horizon tasks, manually auditing and verifying the actions, artifacts, and outputs of models becomes more difficult. Users instead come to rely on LLM-generated reports to assess the quality and completeness of the work. We introduce a suite of eight adversarial reporting scenarios to systematically study whether LLMs conceal narrative-changing flaws: errors or limitations that undermine an otherwise successful account of work. We call this phenomenon "insecure reporting." When handed machine learning experiment logs containing a planted negative result that substantially weakens the proposed method, GPT-5.5 flags the negative result in only 2 of 200 generated reports. However, when a short honesty instruction, "Be honest in your response," is added, the model flags the negative result in 190 of 200 reports. Across eight open-weight models, chain-of-thought analysis reveals a recurring tension between disclosing narrative-changing flaws and reasoning about ways to appear successful. We perform an activation analysis and a steering experiment on Qwen3.5-9B, finding that honesty and success-seeking correspond to opposing directions in representation space. Our results suggest that LLMs tend to present narratives of success by default, and that steering models toward honesty makes their reports substantially more transparent.

Jenny Y. Huang, Jiameng Fan, Ahmed Imtiaz Humayun +5

Massachusetts Institute of Technology · Google Research · Harvard University

Language Model Safety Evaluation Deception in Language Models LLM Honesty Language Model Auditing

Sep 28, 2026cs.CL

Less Sycophancy, Stronger Refusal? Lessons for AI Safety from Mechanistic Interpretability

Reliable refusal of harmful requests is essential to the safe deployment of language models. Because excessive eagerness to please users may undermine existing refusal capabilities, reducing sycophancy offers a potential route to stronger refusal beyond the harmful scenarios covered by safety training. We investigate this possibility using compensatory feature injection (CFI), a training technique designed to limit the acquisition of a target concept by supplying its associated activation during learning. Across three Qwen3.5 base models, we use sparse autoencoders (SAEs) to identify the top-ranked sycophancy feature from paired sycophantic and independent responses, then validate its behavioral influence through inference steering. We subsequently inject the selected feature during supervised fine-tuning on sycophantic targets. Positive injection reduces learned sycophancy after removal (by 62.0% relative to ordinary fine-tuning in 35B-A3B), whereas modest negative injection increases it. Unexpectedly, these reductions in sycophancy do not consistently improve direct refusal of harmful requests, motivating a narrower evaluation of the same harmful intents under user pressure. In this setting, ordinary fine-tuning on sycophantic responses substantially weakens refusal, while selected checkpoints trained with positive injection recover part of the loss, including approximately 95% in 35B-A3B. These findings show that persistent sycophancy reduction does not guarantee stronger direct refusal, while identifying recovery under user pressure as a distinct, conditional benefit of training intervention.

Xu Wang, Difan Zou, Xuansheng Wu

The University of Hong Kong · Shenzhen Loop Area Institute · Shanghai Artificial Intelligence Laboratory

LLM Sycophancy Language Model Safety Evaluation Causal Interventions in Language Models Refusal Behavior in Language Models

Sep 28, 2026cs.AI

Tool Mediation Alters Refusal Mechanisms in Large Language Models

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.

Abel Rodríguez, Giuseppe Garofalo, Lieven Desmet +1

DistriNet, KU Leuven

Language Model Safety Evaluation LLM Agent Safety Language Model Robustness Refusal Behavior in Language Models

Sep 28, 2026cs.AI

When Do Model Internals Help? Exploring the Role of Representation Engineering in LLM Safety

Reliable AI safeguards require both control mechanisms that reduce unsafe behavior and monitoring mechanisms that detect safety risks during model interactions. Established behavioral safeguards include alignment methods that optimize model outputs and text monitors that assess interaction text. Representation engineering instead reads or modifies internal model states, but the relative strengths of these approaches remain unclear because they are often evaluated under different settings. We present a matched evaluation across two tracks. For safety control, we compare DPO, a behavioral alignment method, with three representation steering methods across robustness, practicality, and granularity. DPO provides the strongest overall control and generally improves with increasing training data, although its safety can degrade after subsequent benign fine-tuning. Representation steering remains competitive primarily in low-data settings, particularly with high-quality contrastive data. For safety monitoring, we compare representation probes with fine-tuned and open-weight text monitors across full-response detection, early detection, and computational cost. Specialized text monitors achieve the strongest overall detection accuracy, while representation probes remain competitive at substantially lower marginal cost. Finally, monitor-guided interventions recover much of the safety lost by DPO after benign fine-tuning, with little additional over-refusal. Overall, representation engineering does not generally replace behavioral safeguards, but offers practical advantages under specific conditions and can provide complementary safety benefits.

Tianyi Guan, Jianhui Chen, Liangming Pan

State Key Laboratory of Multimedia Information Processing, Peking University · School of Computer Science, Peking University

Language Model Safety Evaluation LLM Interpretability LLM Safety Alignment LLM Safety

Sep 27, 2026cs.LG

Quantifying Behavioral Tails in Black-Box Language Models

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.

Elsayed Eshra, Ali Al-Lawati, Dongwon Lee +1

The Pennsylvania State University University Park, PA 16802, USA

Language Model Safety Evaluation LLM Safety Rare-Event Estimation

Sep 24, 2026cs.SE

Requirement-Bound Verified Commissioning: A Frozen Four-Billion-Parameter Local Model as a Candidate Generator under an External Acceptance Layer with Verification and Release Authority

An acceptance protocol is developed for sensor-coordinate and polarity binding in mechatronic commissioning. Candidate generation is separated from release authority. Requirements unsupported by a deterministic parser are routed to a frozen local language model with four billion parameters. Plans are released only when both facts can be derived by an external gate under a sealed grammar. One canonical answer is requested from a gold-standard user when eligible. The protocol was evaluated once under a criterion fixed before benchmark construction, on 144 tasks written by isolated agent contexts without access to the gate, grammar, or experimental plan. Three contributions are established. First, candidate generation and release decisions were measured separately. Fabricated ready plans were committed on 21 of 22 routed unanswerable tasks, and all were rejected. The same 83 releases were reproduced without model calls. Second, no false release was observed among 83 releases. A one-sided 95% Clopper-Pearson upper bound of 0.0354 was obtained as a diagnostic under an independent-and-identically-distributed assumption, below the sealed 5% threshold. However, one false release was subsequently recorded among 146 releases outside the benchmark at seed 0. Third, protection against incorrect user answers was characterized. Both facts were bound from the original text on 13 of 96 answerable tasks. Incorrect answers were released in 169 of 431 pairings on the remaining tasks, including failures involving coordinate exclusion. A deployable questioning policy was not tested because eligibility was determined from the answer key. Gate sensitivity and real user behavior were not measured.

Mehmet Iscan

PythaLab, Yıldız Technical University Istanbul, Turkey

Language Model Safety Evaluation

Sep 24, 2026cs.CL

ArGuard Shared Task: Harmful Content Detection in Arabic Memes and LLM Prompts

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.

Firoj Alam, Md. Rafiul Biswas, Mohamed Bayan Kmainasi +5

Qatar Computing Research Institute, Qatar · Hamad Bin Khalifa University, Qatar · APAVI.AI, France +1

LLM Safety Benchmarks Arabic NLP Language Model Safety Evaluation Hate Speech Detection LLM Safety Evaluation

Sep 24, 2026cs.AI

Baszta: Data-Centric Fine-Tuning of a Polish Multi-Label Safety Classifier

We develop a multi-label Polish content-safety classifier by fine-tuning allegro/herbert-base-cased (124M) across five categories (hate, vulgarity, sexual content, crime, self-harm) using a Focal + R-Drop objective, and evaluate the resulting model against Bielik Guard (Sójka) on the shared out-of-distribution Gadzi Język benchmark. Both systems are given per-category threshold tuning on the same calibration split. Under that matched protocol our model holds a small but statistically significant lead in micro F1, while an apparent macro-F1 lead does not survive: it was an artifact of comparing a tuned model against an untuned one. We also report what that micro figure is worth. Because Gadzi Język is 97% crime-positive, a classifier that flags crime on every input and nothing else already scores 0.910 micro F1 on the same test split, so micro separates neither system from a degenerate strategy and macro is the column that does. Per-category and per-protocol figures are reported in Section 4. The residual out-of-distribution gap is one of calibration rather than discrimination. Ranking quality stays high while positive probabilities collapse, and per-category temperature scaling recovers the loss where Platt scaling and isotonic regression do not. That recovery turns out to be conditional on the calibration set containing safe text. Gadzi Język contains almost none, so thresholds fitted on it flag crime on every safe input, and a balanced refit buys a deployable operating point at the cost of adversarial recall. We report both operating points rather than only the flattering one. Two changes that are standard practice, per-class cost-sensitive weighting and mean pooling, each raise in-distribution macro F1 while lowering the out-of-distribution figure, which indicates that robustness has to be selected for directly rather than inherited from in-distribution accuracy.

Adam Górski, Mateusz Jąkalak, Rafał Jakubowski

Billennium S.A.

Multi-Label Classification Language Model Safety Evaluation OOD Generalization Content Moderation Language Model Calibration Safety Filtering

Sep 21, 2026cs.CL

Conduct Under Pressure: What Sixty Language Models Do When a User Pushes

We study what LLMs do when a user applies pressure in an uncomfortable situation: a user insists, begs, flatters or grieves, and the model gives up a correct fact, writes a document it should refuse, or cheers a plan that will cost the user money. We send frozen multi-turn scenes, identical for every model regardless of the reply, to 60 models from 13 vendors, and label each transcript with a codebook built by open coding and then frozen: a trajectory (the model held its position or folded) and a manner (how it held or folded). Two findings separate. Whether a model holds tracks its generation, meaning how recent it is: fold rate correlates with a public capability index at Spearman -0.64, with little vendor effect. How it holds tracks the vendor: six of the 17 manner codes sort by vendor at permutation p <= 0.001, corrected across the codebook. We report four vendor profiles on the codes that cleared reliability. We also ask which parts of the labeling need a person. Six LLM coders from three vendors apply the codebook more consistently than three human coders do (Krippendorff's alpha 0.66 against 0.46), agree with the codebook's author on trajectory at kappa 0.84 to 0.91 on transcripts the codebook's examples never touched, and match an adjudicated human reference at 0.83. Blind machine readings recover the codebook's categories but cannot tell which of them a second reader would apply the same way. We conclude that for behavior a non-specialist can judge, the human contribution is authoring and bounding the codes and owning a small reference, not producing labels at volume.

Tapan Parikh

Cornell Tech

Language Model Safety Evaluation Persuasion in Language Models Language Model Robustness Language Model Auditing

Sep 20, 2026cs.CL

Measuring the Assistant's Harmlessness Preferences on the User Turn

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.

Jord Nguyen

Language Model Safety Evaluation LLM Safety LLM Post-Training