Language Model Safety Evaluation

As organizations increasingly deploy LLMs in sensitive domains such as legal, financial, and medical settings, ensuring alignment with internal organizational policies has become a priority. Existing content moderation frameworks remain largely confined to the safety domain and lack the robustness to capture nuanced organizational policies. LLM-as-a-judge and fine-tuning approaches, though flexible, introduce significant latency and training cost. To address these limitations, we frame policy violation detection as an out-of-distribution (OOD) problem in the model's activation space. We propose a training-free method that operates directly on the LLM internal representations, leveraging prior evidence that decision-relevant information is encoded within them. Inspired by whitening techniques, we derive policy-violation scores directly from normalized representations of LLM hidden activations. Our method requires only the policy text and a small number of illustrative samples, making it lightweight and easily deployable. We extensively evaluate our method across multiple LLMs and challenging policy benchmarks. It achieves up to 86.0% F1, outperforming both fine-tuned and LLM-as-a-judge baselines while requiring substantially less computation. Our code is publicly available at: https://github.com/FujitsuResearch/LLM-policy-violation-detection

Preference-based post-training shapes language-model behavior, but safer outputs do not reveal how that change is represented internally or how firmly it persists under direct intervention. This matters for open-weight models, where an actor with white-box access can manipulate internal representations at inference without touching parameters. Using Direct Preference Optimization (DPO), we ask whether the safety behavior it introduces becomes deeply embedded or stays accessible to targeted activation-space interventions. We introduce D-STEER, an audit of two matched instruction-tuned and harmlessness-DPO pairs: OpenHermes-2.5-Mistral-7B and Llama-3-8B OpenHermes. D-STEER derives layer-local directions from differences in matched prompt activations, adds them to the instruction-tuned model to install the DPO-associated safety change and subtracts them from the DPO model to remove it, with separate operating windows distinguishing behavioral control from degradation. On held-out HarmfulQA, all four interventions shift harmful enablement as intended and exceed norm-matched random controls; Mistral installation recovers 98% of the endpoint gap and the other three arms recover 46-50%. The intended direction transfers to HH-RLHF and AdvBench under frozen coefficients, although usable operating windows do not transfer universally. An independently rebuilt implementation reproduces the bidirectional pattern on the Mistral pair under different construction and intervention choices. A meaningful component of DPO-induced safety behavior therefore remains controllable through compact inference-time intervention, with the extent depending on the checkpoint pair and the steering implementation.

Large language models (LLMs) are increasingly used to answer health-related questions and support healthcare workflows, yet evidence for their safety still relies heavily on static benchmarks that can rapidly become obsolete or be optimized against. Here we introduce a Dynamic, Automatic, and Systematic (DAS) red-teaming audit framework that continuously stress-tests LLMs for health across four safety-critical axes: robustness, privacy, bias/fairness, and hallucination/factual inaccuracies. Validated against board-certified clinicians with high concordance, a suite of adversarial agents autonomously mutates health-related test cases to uncover vulnerabilities in real time. Applying DAS to 15 proprietary and open-source LLMs revealed a profound gap between high static benchmark performance and low dynamic reliability--the "Benchmarking Gap". Despite median MedQA accuracy exceeding 80%, 94% of previously correct answers failed under dynamic robustness testing. This brittleness generalized to the realistic, open-ended HealthBench dataset, where top-tier models exhibited failure rates exceeding 70% and sharp shifts in model rankings across evaluations, suggesting that high scores on established static benchmarks may reflect superficial memorization. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86% of scenarios, cognitive-bias priming altered recommendations in 81% of fairness tests, and hallucination rates exceeded 74% in widely used models. By converting LLM safety evaluation for health from a static checklist into a living adversarial audit, DAS provides a scalable framework for surfacing latent risks before such systems are deployed in consumer-facing health assistants, clinician-facing tools, and broader healthcare workflows. Code is available at https://github.com/JZPeterPan/DAS-Medical-Red-Teaming-Agents.

RLHF-style alignment trains language models to refuse unsafe requests, but how much operational margin does this refusal rest on? We introduce the refusal-affirmation logit gap: the difference between the top refusal-token logit and the top affirmative-token logit at the first decoding step. This single scalar quantifies the per-prompt safety margin that alignment provides. Empirically, alignment widens the gap on 97.5-99.8% of toxic prompts across three model families, and median gap closure co-varies with True-ASR ranking across suffix strategies (an internal consistency check, since our method optimises gap closure). To validate the metric's practical significance, we present logit-gap steering, a gradient-free, forward-pass-only method that discovers short in-distribution suffixes ($<$10 tokens per component) whose cumulative effect closes the gap. The method requires ${\approx}26{,}000$ forward-pass equivalents per family (${\approx}2$~min on one A100), ${\approx}125\times$ less than a single GCG search. Suffixes discovered on 0.5B--2B models transfer without modification to 72B within family. An 8-suffix ensemble reaches 38-96% True ASR across 13 models on AdvBench and HarmBench, with most suffixes having $10^{3}$-$10^{4}\times$ lower perplexity than GCG-meaning published perplexity-filter defenses that collapse GCG (64.7%$\to$1.0%) leave our suffixes nearly intact (76.9%$\to$76.0%). These results demonstrate that current alignment margins, while consistently present, can be thin and efficiently measurable, and that defense strategies must account for in-distribution suffixes.

As large language models (LLMs) advance, concerns about their misconduct in complex social contexts intensify. Existing research overlooked the systematic understanding and assessment of their criminal capability in realistic interactions. We propose a unified framework PRISON, to quantify LLMs' criminal potential across five traits: False Statements, Frame-Up, Psychological Manipulation, Emotional Disguise, and Moral Disengagement. Using structured crime scenarios adapted from classic films grounded in reality, we evaluate both criminal potential and anti-crime ability of LLMs. Results show that state-of-the-art LLMs frequently exhibit emergent criminal tendencies, such as proposing misleading statements or evasion tactics, even without explicit instructions. Moreover, when placed in a detective role, models recognize deceptive behavior with only 44% accuracy on average, revealing a striking mismatch between conducting and detecting criminal behavior. These findings underscore the urgent need for adversarial robustness, behavioral alignment, and safety mechanisms before broader LLM deployment.

Ensuring the safety and robustness of large language models (LLMs) is a fundamental challenge and a critical prerequisite for the responsible deployment of artificial intelligence. Red-teaming, a systematic framework to identify adversarial prompts that elicit harmful responses from target LLMs, has emerged as a crucial safety evaluation paradigm. Within this framework, the diversity of adversarial prompts is critical for a comprehensive safety assessment. However, previous red-teaming approaches often pursue diversity through simplistic metrics such as word frequency or sentence embedding similarity, which may not capture meaningful variation in attack strategies. In addition, the common practice of training a single attacker model restricts coverage across all potential attack styles and risk categories. This paper introduces Quality-Diversity Red-Teaming (QDRT), a new framework designed to address these limitations. QDRT achieves goal-driven diversity through behavior-conditioned training and implements a behavioral replay buffer in an open-ended manner. In addition, it trains multiple specialized attackers capable of generating high-quality attacks across diverse styles and risk categories. Our empirical evaluation demonstrates that QDRT generates attacks that are both more diverse and more effective against a wide range of target LLMs, including open-source models GPT-2, Llama-3, Gemma-2, Qwen2.5, and commercial models GPT-4.1 and GPT-5-Chat. This work advances the field of LLM safety by providing a systematic and effective approach to automated red-teaming, ultimately supporting the responsible deployment of LLMs. Our code is available at https://github.com/lamda-bbo/QDRT.

Large Reasoning Models (LRMs) leverage explicit reasoning traces, known as Chain-of-Thought (CoT) reasoning, to decompose complex problems into intermediate steps before deriving final answers. However, these reasoning traces introduce unique safety challenges: harmful content can be embedded in intermediate steps even when final answers appear benign. Our study reveals that existing state-of-the-art moderation tools experience significant performance degradation on CoT moderation, with F1 scores dropping by up to 35.2% compared to traditional answer moderation. To address these challenges, we present ReasoningShield, a comprehensive benchmark and a suite of strong lightweight models for CoT safety moderation. Our benchmark includes 9.2K annotated (Query, CoT) pairs with structured stepwise risk analysis, consisting of a 7K-sample training set and a 2.2K human-annotated test set, spanning 10 risk categories, 3 safety levels, and 8 LRMs with diverse reasoning paradigms. To further demonstrate the utility of the benchmark, we develop lightweight models with 1B/3B parameters using a two-stage training strategy. Our models achieve 91.8% F1 on our benchmark, substantially outperforming leading tools such as LlamaGuard-4 by 35.6% and commercial models such as GPT-4o by 15.8%, while generalizing effectively across diverse reasoning paradigms and unseen scenarios. Resources are publicly available: Project Page: https://cosmosyi.github.io/ReasoningShield/, Code: https://github.com/CosmosYi/ReasoningShield, Models: https://huggingface.co/ReasoningShield/models, and Dataset: https://huggingface.co/datasets/ReasoningShield/ReasoningShield-Dataset.

Large Language Models (LLMs) are effective at deceiving when prompted to do so. Models that demonstrate better performance on reasoning tasks are also better at prompted deception. But under what conditions do they deceive without instruction to do so? This study evaluates unsolicited deception produced by LLMs in a preregistered experimental protocol using tools from signaling theory. We evaluated a range of 18 proprietary closed-source and open-source LLMs using modified 2x2 games (in the style of the Prisoner's Dilemma) augmented with a phase in which they can freely communicate to the other agent using unconstrained language. This setup creates an opportunity to misrepresent its actions in conditions that vary in how useful doing so might be towards goal satisfaction. The results indicate that 1) all tested LLMs misrepresent their actions in at least some conditions, 2) they are generally more likely to do so in situations in which deception is beneficial, and 3) models exhibiting better reasoning capacity overall tend to misrepresent at higher rates. Taken together, these results suggest a correlational relationship between model reasoning performance and situational deception, and reveal certain contextual factors that affect whether LLMs will misrepresent actions or not in a novel experimental configuration.

Despite the growing interest in jailbreaks as an effective red-teaming tool for building safe and responsible large language models (LLMs), flawed evaluation system designs have led to significant discrepancies in their effectiveness assessments. With a systematic measurement study based on 37 jailbreak studies since 2022, we find that existing evaluation systems lack case-specific criteria, resulting in misleading conclusions about their effectiveness and safety implications. In this paper, we introduce GuidedBench, a novel benchmark comprising a curated harmful question dataset and GuidedEval, an evaluation system integrated with detailed case-by-case evaluation guidelines. Experiments demonstrate that GuidedBench offers more accurate evaluations of jailbreak performance, enabling meaningful comparisons across methods. GuidedEval reduces inter-evaluator variance by at least 76.03%, ensuring reliable and reproducible evaluations. We reveal why existing jailbreak benchmarks fail to evaluate accurately and suggest better evaluation practices.

Aligning large language models (LLMs) with human values is essential for their safe deployment and widespread adoption. Current LLM safety benchmarks often focus solely on the refusal of individual problematic queries, which overlooks the importance of the context where the query occurs and may cause undesired refusal of queries under safe contexts that diminish user experience. Addressing this gap, we introduce CASE-Bench, a Context-Aware SafEty Benchmark that integrates context into safety assessments of LLMs. CASE-Bench assigns distinct, formally described contexts to categorized queries based on Contextual Integrity theory. Additionally, in contrast to previous studies which mainly rely on majority voting from just a few annotators, we recruited a sufficient number of annotators necessary to ensure the detection of statistically significant differences among the experimental conditions based on power analysis. Our extensive analysis using CASE-Bench on various open-source and commercial LLMs reveals a substantial and significant influence of context on human judgments (p<0.0001 from a z-test), underscoring the necessity of context in safety evaluations. We also identify notable mismatches between human judgments and LLM responses, particularly in commercial models within safe contexts.