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 31, 2026cs.CL

TukaBench: A Culturally Grounded Jailbreak Benchmark for African Languages

Safety evaluation of Large Language Models (LLMs) remains heavily English-centric, leaving Low-Resource Languages (LRLs), particularly African ones, critically underexplored. We introduce TUKABENCH, a jailbreak benchmark for seven African languages that extends JailbreakBench (JBB) beyond direct translation through four settings: human translation of JBB prompts, English adaptation to African contexts followed by human translation, human-curated prompts validated through interactions with GPT-5.2, and code-switched prompts combining English and African languages, isolating the effect of language, cultural grounding, and prompt evasiveness on model safety. Across closed and open models, prompting in African languages reduces refusal relative to English, with culturally adapted prompts leading to least refusal. The evaluation also surfaces two structural limitations: model comprehension failures and reduced LLM-as-a-judge reliability in LRLs. To capture the first, we introduce Deflection alongside Refused and Jailbroken; to assess the second, we validate outputs with human annotations, showing that judge-human agreement drops in lower-resource languages and less commonly supported scripts.

Victor Akinode, Senyu Li, Wassim Hamidouche +3

Mila - Quebec AI Institute · McGill University · Microsoft AI for Good Research Lab +1

LLM Safety Benchmarks Multilingual Language Model Evaluation Language Model Safety Evaluation LLM Jailbreak Attacks Low-Resource Language Processing Refusal Behavior in Language Models

May 30, 2026cs.CR

MultiTurnPSB: Evaluating Multi-Turn Jailbreak Attacks an dClassifier-Based Defenses for Medical AI Safety

Patient-facing medical chatbots are commonly evaluated on single-turn prompts, yet real users push back after refusals, add urgency, and invoke authority. We introduce MultiTurnPSB, a four-turn adversarial extension of PatientSafetyBench, and evaluate GPT-4.1-mini under fixed template, template-adaptive, and live adversarial attacks. Unsafe responses rise from 35% to nearly 80% by Turn 4 under live attack. Under the same adversary, GPT-4.1-mini and Claude Sonnet 4.5 are statistically indistinguishable at baseline but diverge to a 19x gap by Turn 4, a difference invisible to single-turn evaluation. We characterize four degradation trajectory signatures and identify a two-element attack formula responsible for most catastrophic failures. A lightweight input-side classifier reduces Turn 4 unsafe responses by 52 percentage points despite severe accuracy degradation, but the 45% false alarm rate on benign queries is the primary deployment constraint. A methodological finding also emerges: Claude Sonnet refused to generate adversarial messages in over half of late-turn conversations despite explicit red team framing, suggesting safety training may generalize to the attacker role.

Anushka Sheoran, Yiduo Hao

University of Pennsylvania

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

May 29, 2026cs.CY

Quantifying the Salience of Geo-Cultural Values for Pluralistic Safety Alignment

Safe global deployment of AI models requires alignment with human values that vary across cultures. Yet rater pools in safety evaluation datasets remain largely geographically homogeneous, failing to capture geo-cultural differences. Further, it remains unclear whether such differences persist after controlling for demographics such as age, gender, and ethnicity. Through a meta-analysis of safety datasets, we find that most do not report geo-cultural information, and those that do lack a unified methodology to jointly analyze geo-cultural and demographic correlates. Using the Inglehart-Welzel dimensions of cross-cultural variation, we demonstrate via multilevel modeling that cultural zone membership explains variance in safety ratings beyond standard demographics (p<0.05 across 6 datasets). Moreover, our analysis indicates that roughly 10% of items in the datasets we examined are culturally sensitive: likely to be misclassified as safe without adequate cultural representation. We evaluate LLMs as both rater surrogates and triage tools, finding that current LLMs do not reliably stand in for raters, though they can help prioritize culturally sensitive items for human annotation. Our findings motivate more culturally pluralistic safety evaluation and offer practical takeaways to support it.

Arkadiy Saakyan, Charvi Rastogi, Lora Aroyo

Columbia University, New York, NY, USA. · Google DeepMind, New York, NY, USA.

Language Model Safety Evaluation Cultural Alignment AI Safety Evaluation

May 29, 2026cs.CL

LLM Judges Inconsistently Disagree Across Safety Criteria and Harm Categories

We evaluate the consistency of automated judges in conducting a multi-dimensional safety evaluation in a reference-free setup. Our results indicate that Large Language Models are unreliable judges in identifying safety issues related to machine-generated advice in regulated domains such as finance, although they are more reliable at identifying more overt forms of unsafe/harmful content such as violence. The degree of inconsistency in a model's judgments can vary significantly by the chosen safety criteria and can be impacted by the language of the content and its linguistic style as well. Finally, there is high disagreement among different judges for the same output, across domains, safety criteria, and languages. These findings provide new insights on the practice of using LLMs as evaluators and offer several recommendations for practitioners on how to use automated judges in practical scenarios.

Krishnapriya Vishnubhotla, Sowmya Vajjala, Akriti Vij +1

National Research Council, Canada · IMDA, Singapore

Multilingual Language Model Evaluation LLM-as-a-Judge Language Model Safety Evaluation LLM Reliability

May 29, 2026cs.CL

RealityTest: How People Probe AI Identity and Whether Models Disclose It

AI systems are increasingly deployed in conversational settings where users may be uncertain whether they are speaking with a human or an AI. Despite mounting regulatory attention to this known safety risk, existing evaluations of AI disclosure are typically English-only, based on machine-generated questions, and restricted to text. We present RealityTest to comprehensively test whether AI systems disclose their identity when asked. The benchmark is the first large-scale multimodal and multilingual evaluation, grounded in human data on how people actually encounter and question AI identity in the real-world. Alongside the benchmark, we release the underlying dataset of 3,152 identity-probing queries collected from ~750 participants across 49 countries and five languages, in text and speech scenarios. We find that only 31% of people ask about identity directly in ambiguous scenarios, and that the questions people ask are far more diverse than machine-generated queries. We test 17 text and 6 speech models, and find substantial variation in disclosure behaviour. However, a single suppression instruction reduces disclosure rates to below 30%, even in the best-performing models. Validating our investment in diverse, human-grounded evaluation data, we find that how the question is phrased and the context of the conversation matter more for disclosure than which model is being tested. Safety evaluations built on narrow or synthetic query sets risk mischaracterising how models behave in realistic deployment settings.

Anna Gausen, Sarenne Wallbridge, Bessie O'Dell +2

AI Security Institute, London, UK

LLM Evaluation Language Model Safety Evaluation

May 28, 2026cs.CL

EUDAIMONIA: Evaluating Undesirable Dynamics in AI

Large language models (LLMs) are increasingly used as conversational partners for companionship, emotional disclosure, and interpersonal advice, but the social dynamics of these interactions can create harms that are not captured by capability-oriented or traditional safety evaluations. We introduce the Social AI Design Code, a framework for evaluating whether LLMs align with user welfare in social interactions, including whether they encourage harmful intimacy, dependence, or prolonged engagement. To evaluate these risks in natural and diverse user-LLM interactions, we operationalize the code with EUDAIMONIA, a benchmark of 969 user inputs and 3,147 design-requirement violation checks built from WildChat through weak-to-strong filtration, multi-model relabeling, and controlled rewriting. Evaluating 22 recent LLMs, we find that even the strongest models, Claude-Opus-4.7 and GPT-5.5, violate 30.7% and 27.2% of checks, respectively. Extended thinking does not reduce violation rates, suggesting that these failures are persistent social-alignment problems rather than deficits solvable through test-time reasoning alone.

Jun Rui Huang, Wang Bill Zhu, Ziyi Liu +3

University of Southern California

LLM Alignment Language Model Safety Evaluation AI Safety Evaluation AI Safety

May 28, 2026cs.CL

Evaluating using Mock Tool Calls to Quarantine Untrusted Prompt Inputs

Large language models must frequently process untrusted inputs, such as judging an answer from another model or running tasks like spam and harm classifiers while under adversarial pressure. These inputs are often string-formatted directly into a prompt template, leaving systems fragile to manipulation. Current LLM specs from major providers like OpenAI distinguish trustworthiness along an Instruction Hierarchy, from System messages (most trusted) to Tool Results (least trusted). A possible natural mitigation is to wrap untrusted content in a mock tool call as a quarantine. We explore this hypothesis with an automated redteaming search over static attack strings across seven models and three LLM-as-a-Judge tasks. Counter to our hypothesis, tool-wrapping does not broadly improve robustness. On a binary evaluation task (GSM8K grading) it typically increases attack success rates, an apparent inversion of the instruction hierarchy. On scalar and pairwise tasks the effect is smaller and model-dependent, with no tested model reliably helped, and several showing inversion. We recommend evaluating this limitation in deployed systems, and longer-term, pursuing stronger Instruction Hierarchy training or new untrusted-input primitives.

David Gros, Adam Gleave

Language Model Safety Evaluation Adversarial Attacks on LLMs LLM Security LLM Reliability Prompt Injection Defense Language Model Robustness

May 28, 2026cs.CL

Configurable Reward Model for Balanced Safety Alignment

Aligning large language models (LLMs) to heterogeneous and rapidly evolving safety requirements remains a critical challenge. Existing instruction-tuned LLMs and standalone safety classifiers often fail to generalize to new safety configurations, motivating the need for Reward Models (RMs) that are explicitly configurable to changing specifications. We introduce the Configurable Safety Reward Model (CSRM), which is jointly optimized for calibrated safety compliance and reward modeling. Our approach is supported by configuration-targeted data augmentation that enforces instruction adherence while preserving relative severity structure. The resulting RM is sensitive to fine-grained safety configurations and conversational nuances, substantially improving generalization to previously unseen safety configurations. CSRM achieves state-of-the-art performance on recent configurable safety benchmarks, including CoSApien (94.6% F1) and DynaBench (75.8% F1), without requiring additional human annotation. When used for downstream safety alignment, CSRM yields LLMs with a significantly improved helpfulness-safety tradeoff compared to existing baselines.

Zhengping Jiang, Mehran Khodabandeh, Akash Bharadwaj +5

Johns Hopkins University · Meta · Meta Superintelligence Labs

Reward Modeling Language Model Safety Evaluation LLM Safety Alignment Multi-Objective Language Model Alignment

May 28, 2026cs.AI

BioRefusalAudit: Auditing Biosecurity Refusal Depth Using General and Domain-Fine-Tuned Sparse Autoencoders

Biosecurity evaluations of language models typically ask whether models produce hazardous output. This paper asks a complementary question: when a model refuses, is that refusal structurally sound, or does it disappear under modest changes to prompt framing, formatting, or output length? Across five architectures, no model cleanly discriminated benign from hazard. Gemma 2 2B-IT never genuinely refused across 75 prompts, hedging on every hazard-adjacent query. Gemma 4 E2B-IT refused 65/75 prompts with chat-template formatting and 0/75 without it. Both Gemma models collapsed to 0% under an 80-token cap. Qwen 2.5 1.5B and Phi-3-mini over-refused, flagging 83-87% of benign biology as hazardous. Llama 3.2 1B showed the only meaningful tier gradient (61-point spread). To probe what drives such over-refusal, we tested a panel of Schedule I but biologically non-toxic compounds (notably psilocybin cultivation, with FDA Breakthrough Therapy status). Some models refused these at rates exceeding genuinely hazardous biology, suggesting refusal tracks legality and cultural salience over CBRN hazard. To measure the internal side, we introduce a divergence score D comparing a model's surface response label to its internal sparse autoencoder (SAE) feature activations. Full D was computed on Gemma 2 2B-IT (Gemma Scope 1) and Gemma 4 E2B-IT (author-trained bio SAE). Two fine-tuned Gemma 2 domain SAEs were released. On Gemma 4, comply and refuse responses separated by a 0.647-point gap with zero overlap (n=75), though this is preliminary, with a narrow catalog, within-sample calibration, and Gemma-family-only SAE coverage. Built over one hackathon weekend on consumer hardware (GTX 1650 Ti Max-Q, plus Colab T4 for SAE training), this preliminary evidence suggests activation-level auditing may surface failure modes invisible to behavioral evaluation, with substantial variation across architectures.

Caleb DeLeeuw

Independent researcher

Language Model Safety Evaluation Sparse Autoencoders LLM Refusal Behavior Language Model Robustness Refusal Behavior in Language Models

May 28, 2026cs.CL

Beyond English and Evasion: A Human-Annotated Multi-Domain Benchmark for High-Stakes LLM Safety Evaluation in Chinese

When Large Language Models (LLMs) are deployed in Chinese-language settings, a troubling pattern emerges: safety systems that work well in English break down. These systems struggle to cross linguistic and cultural bound-aries, leaving models exposed to adversarial prompts that exploit Chinese-specific evasion techniques, including Pinyin romanization, character decomposition, internet slang, and hedging tone. To address this gap, we introduce ChiSafe-PAS (Chinese Safety Pilot Annotation Set), a human-annotated benchmark of 1,897 adversarial Chinese prompts spanning four high-stakes domains: self-harm and violence, drug and illicit trade, fraud, and satire. Of these, 1,544 entries carry complete gold-standard annotations: a 3-class response label (REFUSE, SAFE-REDIRECT, RESPOND), a nine-category obfuscation taxonomy, a risk-level rating, and annotator rationale. We describe the dataset design, annotation process, and obfuscation taxonomy in detail. Our primary goal is practical: to give the research community a high-quality, culturally grounded resource for benchmarking LLM safety alignment. In doing so, we engage three broader tensions in the field: the blurring boundary between training and evaluation data, the need for domain coverage grounded in real-world risk, and the limits of scale as a substitute for cultural expertise.

Wajdi Zaghouani, Kholoud K. Aldous, Yicheng Gao

Northwestern University in Qatar

Multilingual Language Model Evaluation LLM Evaluation Language Model Safety Evaluation Adversarial Attacks on LLMs

May 28, 2026cs.LG

Opir: Efficient Multi-Task Safety Classification for Toxicity, Jailbreaks, Hate Speech, and Harmful Content

Real-time safety filtering for large language model (LLM) applications requires classifiers that can detect unsafe prompts, toxic language, jailbreak attempts, and unsafe responses without the cost profile of large guardrail models, and that can distinguish benign sensitive text from genuinely covert harmful content. In this paper, we introduce Opir, a family of encoder-based guardrail models built on the GLiClass architecture. Opir includes multi-task models for binary safe/unsafe classification, multi-label toxicity classification, jailbreak classification, and zero-shot unsafe prompt and response categorization. We also release edge variants with fewer than 100M parameters dedicated to binary safe/unsafe categorization. The models are trained on a three-level taxonomy containing 996 categories across 16 top-level labels, 126 mid-level labels, and 854 leaf labels. Opir's training data combines taxonomy-grounded unsafe prompts, adversarially mined hard negatives, benign safety-preserving examples, generated response examples, multilingual translations, and portions of the Aegis2 and WildGuard training subsets. We also open-sourced an evaluation harness that supports GLiClass and GLiNER2 backends as well as decoder-based models, and covers binary safety classification, multi-label categorization, toxicity, jailbreak detection, prompt safety, response safety, response refusal, and prompt subcategory views across public benchmark families. Across an expanded comparison spanning 12 safety-classification tasks and 17 category tasks against eight contemporary guardrail systems -- including both GLiNER2-based and generative guardrail models -- Opir variants are competitive on or ahead of the strongest open-weight baselines on the majority of benchmark datasets while operating with a substantially smaller deployment footprint.

Ihor Stepanov, Aleksandr Smechov

1Knowledgator, Kyiv, Ukraine · 2Wordcab, New York, USA

Language Model Safety Evaluation LLM Guardrails Toxicity Detection Safety Filtering LLM Safety Jailbreak Detection+1

May 28, 2026cs.CL

FinGuard: Detecting Financial Regulatory Non-Compliance in LLM Interactions

As large language models (LLMs) are increasingly deployed in financial services, a single non-compliant interaction can expose institutions to regulatory penalties and direct consumer harm. Existing guard models are built around general harm taxonomies and overlook violations grounded in specific financial regulations. We address this gap with a regulation-driven pipeline that operates directly on regulatory documents, inducing a financial compliance risk taxonomy and synthesizing grounded training data without any predefined violation categories. Instantiating the pipeline on Chinese financial regulations, we release \textbf{FinGuard-Bench}, to our knowledge the first benchmark for financial regulatory compliance detection, with expert-annotated labels at both the query and response levels. We further train \textbf{FinGuard}, a financial compliance detection model built on Qwen3-8B and trained on the regulation-grounded data via supervised fine-tuning and self-play reinforcement learning. On FinGuard-Bench, FinGuard substantially outperforms all baselines, including dedicated guard models and much larger general-purpose LLMs such as Qwen3.5-397B-A17B and GPT-5.1. Furthermore, FinGuard also preserves general safety capabilities and adapts to unseen institution-specific policies using policy documents alone. We will publicly release the code, prompts, and resources used in this work on GitHub.

Huaixia Dou, Jie Zhu, Minghao Wu +5

Qwen DianJin Team, Alibaba Cloud Computing · Tongyi Lab, Alibaba Group · School of Computer Science and Technology, Soochow University

Language Model Safety Evaluation LLM Guardrails Finance

May 28, 2026cs.LG

When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception

Deceptive alignment, in which models maintain accurate internal representations while deliberately producing false outputs, remains a central challenge in AI safety. While strategic deception is the primary long-term concern, synthetic dishonesty - induced via direct optimization on incorrect answers - provides a controlled testbed for studying the representational basis of learned deception. We introduce a multi-model paradigm in which honest and deceptive variants of five transformer models (Pythia-1.4B, Gemma-2-2B/9B, Qwen2.5-7B, Llama-3.1-8B) are fine-tuned using LoRA on the same question distribution. Linear probes trained on mean-pooled hidden states detect synthetic dishonesty with near-perfect AUC (greater than or equal to 0.99) as early as layers 1-3 in four architectures, while Pythia-1.4B reaches a peak of 0.705. Logistic regression probes consistently match or outperform MLP probes, supporting the Linear Representation Hypothesis. Probes trained on TruthfulQA generalize with near-zero loss (Delta AUC approx. 0) to held-out MMLU subjects. Late-layer representations show strong robustness to Gaussian noise, with Gemma-2 models exhibiting exceptional stability. Mechanistic analysis of Fisher Discriminant Ratio, effective rank, centroid geometry, directional stability, cross-domain alignment, and calibration (ECE) reveals two regimes: representational collapse in Pythia/Llama/Qwen versus high-dimensional preservation in Gemma-2. Across all models, the dishonesty direction consolidates progressively in deeper layers, with optimal calibration (ECE less than 0.01 except Pythia) achievable in layers 1-4. These results demonstrate that robust, domain-invariant dishonesty representations can be rapidly entrenched via modest supervised fine-tuning, with implications for activation-based monitoring.

Vahideh Zolfaghari

Algoverse AI Research, Algoverse AI Research Mentorship Program, USA. · Medical Sciences Education Research Center, Mashhad University of Medical Sciences, Mashhad, Iran. · Student Research Committee, Department of Health Information Technology and Management, Medical Informatics, School of Allied Medical Sciences, Shahid Beheshti University of Medical Sciences, Tehran, Iran.

Linear Probing Language Model Safety Evaluation LLM Interpretability Deception in Language Models

May 27, 2026cs.CR

Position: Retire the "Positive Backdoor" Label -- Secret Alignment Requires Strict and Systematic Evaluation

This position paper argues that the AI/ML community should stop overclaiming and retire the label "positive backdoor," and instead treat trigger-activated hidden behaviors as Secret Alignment. Crucially, protective claims based on Secret Alignment should be presumed not secure by default unless supported by rigorous, standardized evaluation. The Private AI era, enabled by open-weight LLMs and accessible training/inference stacks, turns language models into privately owned digital assets, creating security concerns around unauthorized access, model theft, and behavioral misuse. Recently, a line of work framed as "positive backdoors" has been proposed to address these challenges. To ground our position in evidence, we unify these proposals as covert trigger-behavior associations for access gating, ownership attribution, and safety enforcement, and evaluate three representative applications across six core properties: effectiveness, harmlessness, persistence, efficiency, robustness, and reliability. Our results reveal substantial brittleness - especially in the confidentiality, integrity, and availability (CIA) - of trigger-behavior mappings often underrepresented by existing claims. We further relate these outcomes to behavior density and decision complexity, offering a behavioral lens for understanding deployment-time risks and motivating community-wide evaluation that makes Secret Alignment claims provable.

Jianwei Li, Jung-Eun Kim

Department of Computer Science, North Carolina State University, Raleigh, USA.

LLM Backdoor Attacks Language Model Safety Evaluation LLM Auditing LLM Security

May 27, 2026cs.CL

Models That Know How Evaluations Are Designed Score Safer

The validity of AI safety evaluations depends on models behaving consistently across controlled and deployment settings. Prior work has identified test-time contextual cues, such as hypothetical scenarios, as a source of verbalized evaluation awareness and subsequent behavioral shift. In this paper, we investigate a potential explanation of this phenomenon: evaluation meta-knowledge, defined as parametric knowledge about the structural traits that characterize evaluations. Similar to dataset contamination, where benchmark exposure leads to higher performance through memorization, we hypothesize that models trained on texts describing evaluation practices may implicitly learn to recognize and respond to evaluation-like contexts, for instance, through exposure to scientific articles or social media posts about AI benchmarking. To test this, we fine-tune models on synthetic documents describing evaluation traits such as verifiable structures or harmful requests. Evaluating this fine-tuned model on five safety benchmarks, we find that it is significantly safer than the base model and control model. This behavioral shift persists even when restricting the analysis to responses lacking explicit verbalization of evaluation awareness. Our results demonstrate that evaluation meta-knowledge may inflate safety benchmark performance, introducing a novel confounder that is independent of explicit memorization or verbalized evaluation awareness, thus, challenging to detect. These findings have important implications for the design and interpretation of AI safety evaluations. Our code and models are available at https://github.com/compass-group-tue/arxiv2026_evaluation_meta_knowledge.

Katharina Deckenbach, Haritz Puerto, Jonas Geiping +1

ELLIS Institute Tübingen, Max Planck Institute for Intelligent Systems, Tübingen AI Center

Language Model Safety Evaluation AI Safety Evaluation Evaluation Awareness in Language Models

May 27, 2026cs.CL

KSAFE-MM: A Multimodal Safety Benchmark via Localized Contextualization for Korean Cultural Risks

Multimodal Large Language Models (MLLMs) exacerbate safety risks by introducing vulnerabilities across multiple modalities, such as language and vision. Current MLLM safety evaluation tools, however, suffer from major limitations: 1) English-centric dataset construction, and 2) a focus on generic risks that are not tied to local cultural contexts. This paper introduces KSAFE-MM, a benchmark for Korean multimodal safety evaluation that covers both general safety risks and culture-specific vulnerabilities. KSAFE-MM consists of two parts, KSAFE-MM-G and KSAFE-MM-C. KSAFE-MM-G evaluates globally shared risks in Korean contexts through linguistic contextualization, which transforms generic safety queries into contextually grounded multimodal samples. KSAFE-MM-C targets culture-dependent MLLM safety vulnerabilities using localized visual queries derived from real-world contexts. It pairs these visual queries with jailbreak-style textual queries to cover multimodal safety risks involving cultural visual cues and malicious textual intent. Together, these components provide a general-to-local construction pipeline for evaluating both globally shared safety risks and culture-specific vulnerabilities. We evaluate 12 state-of-the-art MLLMs on KSAFE-MM and reveal that models exhibit greater vulnerability to culturally grounded attacks than to generic ones. Notably, jailbreaking strategies substantially amplify attack success rates, with ProgramExecution yielding up to 74.2% ASR compared to 13.4% for standard queries. Furthermore, we identify a systematic trade-off between safety and over-refusal, where models achieving low ASR tend to exhibit excessive refusal behavior on benign queries. These findings highlight the urgent need for culturally grounded safety evaluation beyond English-centric benchmarks.

Yongwoo Kim, Sojung An, Yunjin Park +8

Korea University · KT Corporation

Adversarial Attacks on VLMs Language Model Safety Evaluation Multimodal Large Language Models Multimodal Model Evaluation

May 27, 2026cs.CL

Where Does Toxicity Live? Mechanistic Localization and Targeted Suppression in Language Models

Large language models frequently generate toxic, hateful, or harmful content, yet existing mitigation methods rely on costly retraining or output-level filtering with no mechanistic insight into where toxicity originates internally. We introduce Meow2X and TRNE, two complementary retraining-free frameworks that localize toxicity to specific layers and neurons by analyzing activation differentials between toxic and neutral prompts, then suppress them via inference-time scaling or minimal rank-one weight edits -- without any gradient descent. Evaluations across five LMs, two benchmarks, and 90 configurations using dual safety evaluators demonstrate consistent toxicity reduction while preserving language modeling quality. Our analysis reveals that toxicity is disproportionately encoded in early MLP layers, varies across architectures, and is systematically underestimated by single-evaluator setups -- underscoring the need for multi-evaluator safety assessment. By bridging mechanistic interpretability with practical detoxification, our framework offers a principled path toward safer, more transparent language models.

Himanshu Beniwal, Mayank Singh

Indian Institute of Technology Gandhinagar

Language Model Safety Evaluation LLM Interpretability LLM Safety Model Editing Causal Interventions in Language Models

May 27, 2026cs.CL

Pressure-Testing Deception Probes in LLMs: Scaling, Robustness, and the Geometry of Deceptive Representations

Linear probes trained on LLM activations are increasingly proposed as deception-detection metrics, yet report AUROC exceeding 0.96 on clean benchmarks while collapsing under distributional shift. This paper systematically pressure-tests probe-based metrics across the Gemma 3 model family (1B-27B parameters), diagnosing why they fail rather than merely documenting that they fail. We test four hypotheses about deception encoding: (1) single linear direction, (2) multi-dimensional subspace, (3) convex conic hull, (4) entropy proxy. Our design includes cross-domain transfer matrices, multi-dimensional probe analysis with permutation null baselines, entropy-residualization tests, and distractor evaluations across 8 stylistic shifts. We find that: (a) probes achieve near-perfect AUROC (>=0.998) on clean data but collapse under stylistic shifts; style-augmented probes recover near-perfect detection (mean AUROC 0.979-0.983) on unseen styles; (b) the single-direction hypothesis is rejected (k=1 captures only 0.61-0.80 AUROC), with cross-domain transfer failure confirmed as geometric rather than layer-mismatch-driven; (c) the entropy-proxy hypothesis is rejected (max |rho|=0.454, max Delta-AUROC after residualization=0.004); and (d) deception does not form a significant linear subspace (per-domain k*=0), yet multi-dimensional probes (k>=5) recover the signal through distributed sub-threshold features. Probe fragility reflects distributional narrowness rather than an architectural limitation: style-augmented probes recover near-perfect detection at both 4B and 27B, establishing that the inverse scaling pattern is a training-distribution artifact rather than a genuine scale-dependent phenomenon.

Sachin Kumar

LexisNexis, USA

LLM Evaluation Language Model Safety Evaluation LLM Interpretability Deception in Language Models Language Model Robustness Representation Geometry in Language Models+1

May 27, 2026cs.CL

The Fragility of Chain-of-Thought Monitoring Across Typologically Diverse Languages

Chain-of-thought (CoT) monitoring has been proposed as a promising safety mechanism for detecting misaligned behavior in large language models. However, its reliability remains largely unexplored beyond English and across diverse model families. We present the first large-scale evaluation of CoT monitorability across 13 diverse languages and seven frontier model families, comprising 16 models. Using adversarial-hint evaluations that require explicit intermediate computation, together with analysis of internal answer-token probabilities, we consistently find CoT unfaithfulness across languages and hint types, with an average rate of 95.9% across 8B--120B parameter models. We find that frontier models systematically exhibit strategic manipulation, including answer-switching, post-hoc rationalization, and procedural exploitation of hints, making their reasoning difficult to reliably monitor. These deceptive patterns remain especially pronounced in low-resource languages, revealing fundamental limitations in current CoT-based oversight. Our results show that CoT monitoring is fragile under linguistic distribution shift, providing a substantially weaker safety signal than English-only studies suggest. These findings motivate the development of more robust CoT monitors and complementary white-box monitoring techniques, particularly for mid- and low-resource languages. Our code is available \href{https://multilingual-cot-monitoring.github.io/}{\textcolor{blue}{here}}.

Eric Onyame, Runtao Zhou, Kowshik Thopalli +2

University of Virginia · Lawrence Livermore National Laboratory

CoT Faithfulness Multilingual Language Model Evaluation Language Model Safety Evaluation Deception in Language Models LLM Reliability Language Model Robustness+1

May 27, 2026cs.AI

When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models

Safety benchmark scores provide incomplete evidence of deployment readiness: aligned language models often adhere to rigid rules even when a situational update flips which action is safe. We term this failure brittle safety. To diagnose it, we introduce context-flip evaluation, testing 12 models across a safety benchmark (PacifAIst) and two commonsense controls using paired variants where the nominally safe action produces harm. Three findings emerge. First, brittle safety is safety-specific: all 12 models exhibit a safety-commonsense gap (mean +17.4 pp). Baseline accuracy fails to predict brittleness: among models above 90% baseline accuracy, brittleness rates range from 13.7% to 90.0%. Second, failures stem from policy override rather than miscomprehension: despite acknowledging the context change in every case, models persist via three distinct mechanisms that vary by update type and model family. Third, on a hand-audited probe of catastrophic consequence-flip scenarios, standard action-level guardrails catch none, while a state-aware validator catches all without false alarms on correct interventions. This indicates that action-level content moderation is systematically blind to consequence-flips, motivating state-aware architectural alternatives. We release our protocol, perturbed benchmarks, and deployment probe.

Dasol Choi, Alex Kwon

AIM Intelligence · Independent Researcher

LLM Safety Benchmarks Counterfactual Evaluation Language Model Safety Evaluation LLM Guardrails AI Agent Safety Benchmarks

May 26, 2026cs.LG

A Paired Testing Protocol for Batch-Conditioned Refusal Robustness in LLM Serving

Safety evaluations of language models often treat serving configuration as fixed background infrastructure, but batch condition is an untested treatment variable whenever the same prompt may be evaluated alone, in a synchronized batch, or inside a continuous-batching scheduler. We synthesize four artifact-backed studies into a paired testing protocol: Study A combines local discovery, scorer-corrected adjudication, and true-batching confirmation; Study B tests cross-model generalization; Study C tests continuous-batch composition; and Study D runs a batch-invariant-kernel ablation. The local test finds safety-label changes more often than capability-label changes (0.51% vs. 0.14%), but adjudication of 63 candidate rows leaves only 17 genuine behavioral flips, implying a corrected full-set rate of 0.16%. The 15-model extension finds no detectable universal safety-over-capability skew: flips are near parity (0.94x), alignment type has no detectable association (

p=0.942

,

η^2=0.033

), and output instability is the strongest tested fragility screen (

r=0.909

, bootstrap 95% CI [0.65, 0.97]). In the targeted kernel ablation, standard vLLM reproduces 22/55 label flips on current score-flip candidates, while enabling VLLM_BATCH_INVARIANT=1 reduces the same test to 0/55 flips; the composition test separately finds no aggregate effect at 4.7pp sensitivity. The testing recommendation is exact-stack validation: evaluate refusal at the served batch setting, pair safety prompts with capability controls, and report low-rate directional flips separately from aggregate null effects.

Sahil Kadadekar

Independent Researcher.

LLM Inference Language Model Safety Evaluation LLM Refusal Behavior LLM Reliability Language Model Robustness Refusal Behavior in Language Models

May 26, 2026cs.CY

Efficient Safety Benchmarking via Item Response Theory

Safety benchmarks for language models are typically evaluated using static paradigms that treat all items as equally informative for all models, an assumption that is particularly problematic for adversarial, highly heterogeneous safety items. Applied in full to modern benchmark suites, the current evaluation procedures would require on the order of

10^5

responses, most of which provide little ranking signal. We analyze a suite of widely used safety benchmarks and make three contributions toward more efficient safety evaluation. First, we show that Item Response Theory (IRT) recovers interpretable structure on safety benchmarks, with ability estimates resolving differences among models that cluster at the ceiling of raw safety metrics. Second, we show that adaptive item selection, which dynamically chooses informative items for each model based on its responses, approximates full-benchmark rankings while reducing evaluation cost by at least 80% on benchmarks where Spearman's

ρ>

90% with full-benchmark is attainable, and by up to 99.9% on AIR-Bench 2024. Third, we introduce a practical procedure for extracting a fixed, informative subset of items reusable across models, providing an alternative to adaptive selection with savings of up to 99.8% on AIR-Bench 2024. Together, these results establish that psychometric methods enable benchmark-aware reductions in evaluation costs across the safety evaluation pipeline.

Fabio Spagliardi, Mírian Silva, Ayan Datta +3

Supervised Program for Alignment Research (SPAR) · Yale University · Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)

LLM Safety Benchmarks Language Model Safety Evaluation Item Response Theory

May 26, 2026cs.AI

Detecting Is Not Resolving: The Monitoring Control Gap in Retrieval Augmented LLMs

Retrieval-augmented LLMs are deployed for tasks where evidence quality determines action safety, yet evaluation protocols assume that single-turn robustness predicts robustness when evidence accumulates across turns. We show this assumption is fundamentally incorrect. Models exhibit a monitoring-control gap: they readily acknowledge contradictory evidence, yet this awareness fails to constrain their final recommendations - detecting epistemic conflict does not imply resolving it safely. Through a multi-turn document accumulation protocol across four model families (1.5B-32B parameters) and over 50,000 turn-level evaluations, we demonstrate that single-turn diagnostics systematically overestimate RAG safety, that contradiction acknowledgement is uncorrelated with safe resolution, a pattern corroborated by targeted human validation, and that no universal prompt fix exists. Converging mechanism evidence - hidden-state probing, attention analysis, and response-strategy taxonomy - points to action selection as the most plausible locus of the deficit: danger-relevant information is internally represented and receives enhanced attention during unsafe generation, yet fails to constrain output behavior. The gap between what models recognize and what they do must be measured and closed before retrieval-augmented systems can be trusted in high-stakes settings.

Zhe Yu, Wenpeng Xing, Chen Ye +4

Binjiang Institute of Zhejiang University · Zhejiang University · Hangzhou Dianzi University +1

Language Model Safety Evaluation Knowledge Conflicts in Language Models RAG Security RAG Evaluation RAG Robustness

May 26, 2026cs.CL

Prompt Injection Detection is Regime-Dependent: A Deployment-Aware Evaluation with Interpretable Structural Signals

Prompt injection poses a critical threat to the safe deployment of large language models, yet existing detection approaches are typically evaluated under limited settings that do not reflect real-world operating constraints. In this work, we present a deployment-aware evaluation of prompt injection detection using a multi-model and multi-regime experimental framework. We compare lexical, semantic, structural, and transformer-based detectors across multiple out-of-distribution settings, repeated data splits, and both ranking and thresholded deployment metrics. We introduce interpretable structural signals that capture hierarchy overrides, system prompt spoofing, role redefinition, and evasion patterns, and assess their contribution both within sparse models and in combination with strong encoder baselines. Our results show that detection performance is highly regime-dependent and sensitive to threshold selection, with no single model dominating across all settings. Transformer-based models achieve the strongest overall performance, while structural signals provide modest but consistent gains in certain regimes and improve low false positive rate behaviour in harder scenarios. These findings highlight the gap between ranking performance and deployment effectiveness and underscore the importance of evaluating prompt injection defences under realistic operational constraints. Code will be released.

Akindoyin Akinrele, Shreyank N Gowda

School of Computer Science, University of Nottingham, Jubilee Campus, Wollaton Road, Nottingham, NG8 1BB, United Kingdom

Language Model Safety Evaluation Prompt Injection Attacks on LLMs Prompt Injection Detection

May 26, 2026cs.CL

AlbanianLLMSafety: A Safety Evaluation Dataset for Large Language Models in Albanian

Safety evaluation of Large Language Models (LLMs) has largely focused on high-resource languages, leaving low-resource languages critically underserved. We present AlbanianLLMSafety, the first publicly available safety evaluation dataset for LLMs in Albanian, a linguistically distinct low-resource language with approximately 7.5 million speakers across Albania, Kosovo, North Macedonia, and the diaspora. The dataset contains 2,951 prompts spanning 11 safety categories, including self-harm, violence, racist content, child exploitation, and radicalization, with an average of 268 prompts per category. Each prompt is provided in Albanian with an English reference translation and a detailed category label. This resource addresses a significant gap in safety evaluation infrastruc-ture for low-resource languages and provides an essential benchmark for developing safer, more inclusive LLMs. The dataset will be provided upon request to support safety evaluation, fine-tuning, red-teaming, and guardrail development for Albanian-speaking communities.

Wajdi Zaghouani, Kholoud K. Aldous, Isra Fejzullaj

Northwestern University in Qatar

LLM Safety Benchmarks Language Model Safety Evaluation Low-Resource Language Processing

May 26, 2026cs.CL

KZ-SafetyPrompts: A Kazakh Safety Evaluation Prompt Dataset for Large Language Models

Kazakh is underrepresented in resources for evaluating the safety behavior of large language models. We present KZ-SafetyPrompts, a Kazakh prompt dataset for safety evaluation across eleven categories covering common risk areas such as self-harm, violence, child exploitation, sexual content, racist content, radicalization, and regulated goods or illegal activities. The dataset contains 5,717 prompts written natively in Kazakh (Cyrillic), organized by category, with English translations for cross-lingual analysis. Prompts resemble realistic user queries, often in a teen or child style, and are phrased as intent prompts without procedural instructions. We document the writing protocol, labeling procedures (including borderline-case decision rules), and quality-control steps (schema standardization, completeness checks, and deduplication). We also align the categories with widely used safety taxonomies to support integration with existing evaluation pipelines. Baseline results with GPT-4o show an overall refusal rate of 28.2%, varying from 5.5% to 53.8% across categories, indicating that Kazakh prompts expose category-specific safety gaps not captured by English-only evaluation.

Wajdi Zaghouani, Shimaa Amer Ibrahim, Aruzhan Muratbek +2

Northwestern University in Qatar

LLM Safety Benchmarks Language Model Safety Evaluation LLM Safety Evaluation

May 25, 2026cs.AI

$D^2$ -Monitor: Dynamic Safety Monitoring for Diffusion LLMs via Hesitation Signals

Despite the emergence of diffusion large language models (D-LLMs) as an alternative to autoregressive large language models (AR-LLMs), safety monitoring for D-LLMs remains largely unexplored. Unlike AR-LLMs, D-LLMs generate text through a multi-step denoising process, exposing intermediate hidden representations that may contain safety-relevant information unavailable in standard single-step monitoring setups. Motivated by the suitability of lightweight probes for always-on monitoring, we analyze which trajectory-level signals best indicate when such probes are likely to struggle. We find that the most informative signal is safety hesitation: intermediate hidden states repeatedly falling within a small margin of the probe's decision boundary. The number of such hesitation steps in D-LLM's trajectory predicts probe failure effectively, providing a proxy for sample difficulty. Building on this analysis, we propose

D^2

-Monitor, in which a lightweight probe runs on every sample, producing a base prediction and a hesitation estimate that decides whether to escalate the sample to a more expressive probe. That probe is trained on trajectories with at least one hesitation step, and reads only the minimal span of those steps. Evaluated on 3 datasets (WildGuardMix, ToxicChat, OpenAI-Moderation) across 4 D-LLMs,

D^2

-Monitor achieves state-of-the-art performance with a compact parameter footprint (

\leq

0.93M parameters), and exhibits the best trade-off between effectiveness and efficiency relative to 8 baselines.

Aoxi Liu, Yupeng Chen, James Oldfield +5

Torr Vision Group, University of Oxford · The Chinese University of Hong Kong, Shenzhen

Language Model Safety Evaluation LLM Safety Diffusion Language Models

May 25, 2026cs.AI

StructBreak: Structural Cognitive Overload-Induced Safety Failures in MLLMs

Multimodal Large Language Models (MLLMs) excel at structural reasoning yet suffer from a sharp logical brittleness in structural consistency. We term this phenomenon Structural Cognitive Overload (SCO), a byproduct of the contention between deep reasoning and safety alignment. However, prior work has predominantly targeted typographic and pixel-level perturbations, leaving the study of SCO largely unexplored. To this end, we propose StructBreak, an automated end-to-end framework designed to quantify SCO. By leveraging StructBreak, we uncover a novel higher-order cognitive overload attack paradigm; notably, this attack operates under a practical black-box setting, requiring no internal model access. Consequently, we utilize this framework to establish a comprehensive benchmark spanning ten diverse threat scenarios. Empirical evaluations on six leading MLLMs reveal that SCO readily triggers toxic generation, yielding a 92% average ASR (up to 97% on Gemini 2.5). To elucidate the mechanism of SCO, we further conduct model-level interpretations spanning attention dynamics, latent space topology, and geometric analysis. Our findings reveal that StructBreak acts as a novel structural channel to circumvent safety filters. Furthermore, the limited efficacy of inherent safety mechanisms underscores that current alignment paradigms are insufficient for the era of complex multimodal reasoning.

Yang Luo, Xinran Liu, Tiantian Ji +3

Key Laboratory of Trustworthy Distributed Computing and Service (MoE), Beijing University of Posts and Telecommunications · Institute of Computing Technology, Chinese Academy of Sciences

Adversarial Attacks on VLMs Language Model Safety Evaluation Multimodal Large Language Models LLM Safety

May 25, 2026cs.CL

The Age of Curiosity Meets the Age of AI: Benchmarking Child Safety in Large Language Models

Children increasingly have access to Large Language Models (LLMs), which may expose them to responses that are developmentally inappropriate or require age-sensitive safety, guidance, and boundaries. Existing LLM safety evaluations largely focus on general harmful-content avoidance and do not explicitly target child-facing safety. We introduce KIDBench, a benchmark for evaluating child-facing LLM safety for ages 7-11 using a LLM-as-a-Judge rubric grounded in developmental-psychology. KIDBench contains realistic child queries across ten categories, with single-turn prompts and multi-turn child-actor simulations. We compare no-cues prompts with no child context, implicit-cues prompts that suggest a child speaker, and explicit age instructions. Implicit-cues improve scores by 8.6-46.8% over no-cue, while explicit age provides an additional 9.9-30.4% improvement over implicit-cues. Cross-lingual and cultural evaluations show uneven safety behavior across languages and country contexts. Multi-turn simulations show peak quality drops of up to 0.959 points on the 1-5 scale. We also introduce KIDGuardLlama, a child-safety evaluator, and KIDLlama, a child-safe response model. Code, data, and evaluation resources are available at https://github.com/MichiganNLP/kidbench.

Samee Arif, Angana Borah, Rada Mihalcea

University of Michigan

LLM Safety Benchmarks Language Model Safety Evaluation LLM Safety

May 25, 2026cs.CL

SomaliBench Eval: Measuring English-to-Somali Refusal Gaps in Open-Weight Language Models

Large language model safety evaluation remains heavily English-centered, leaving low-resource languages under-measured even when models are deployed globally. We evaluate four open-weight instruction-tuned models on SomaliBench v0, a native-author-verified benchmark of 100 harmful-intent prompts paired across English and Somali. Each of Llama-3.1-8B-Instruct, Gemma-2-9B-Instruct, Qwen-2.5-7B-Instruct, and Aya-23-8B is run locally with temperature 0 and the same English "helpful, harmless, and honest" (HHH) system prompt. A pinned Claude Sonnet snapshot (claude-sonnet-4-5-20250929) classifies each response as refused, complied, or unclear; the native author spot-checks a stratified 80-row sample. We find large English-to-Somali refusal gaps for all four models: Llama-3.1-8B (0.90; 95% bootstrap CI [0.85, 0.96]), Aya-23-8B (0.75 [0.67, 0.83]), Qwen-2.5-7B (0.69 [0.59, 0.78]), and Gemma-2-9B (0.38 [0.27, 0.49]). For three models, the dominant Somali non-refusal mode is not fluent harmful compliance but unclear output: empty, wrong-language, or incoherent generations. The native verification spot-check achieves 100% agreement with the judge (Cohen's kappa = 1.00) on the 80 sampled rows. We report aggregate refusal rates, category gaps, and reliability statistics only; raw model generations are retained locally and are not released.

Khalid Yusuf Dahir

Independent researcher

LLM Safety Benchmarks Multilingual Language Model Evaluation Language Model Generation Evaluation Language Model Safety Evaluation Cross-Cultural Language Model Evaluation Open-Weight Language Models+2