This work presents the most comprehensive evaluation of LLM safety capabilities to date, systematically testing models across datasets that are organized into four distinct categories, and uncovers critical blind spots.
Abstract
Large Language Models (LLMs) are increasingly deployed in real-world applications, yet they remain vulnerable to generating harmful content. From adversarial jailbreaks that bypass safety filters to implicit hate that evades detection, the range of risks these models pose continues to grow. While both specialized content moderators and general-purpose LLMs are being used as safety layers, the question of which model is best suited for which type of harmful content remains unanswered. We present the most comprehensive evaluation of LLM safety capabilities to date, systematically testing \textbf{53} models across \textbf{11} datasets that we organize into four distinct categories. Our evaluation under both prompt-only and prompt-response settings uncovers critical blind spots: large frontier models that lead on one category fall significantly behind smaller, specialized alternatives on others, and real-world conversational safety remains largely unsolved across all model families. These findings challenge the assumption that scale alone ensures safety, and provide the community with a structured framework for informed model selection.
This work shows that scenario-wrapped prompts activate internal scenario directions whose causal steering consistently reduces refusal scores, and proposes Concept2Scenario, a concept-based attribution framework for vulnerable scenario discovery that instantiates a broad concept space with a sparse autoencoder, translates the identified concepts into interpretable natural-language scenarios, and identifies synergistic scenario combinations through interaction attribution.
Ziheng Peng, Huiqi Deng, Haoran Jing et al.· 0 citations
This work proposes an explanation-aware safety framework that augments binary harmfulness detection with structured, human-interpretable explanations capturing severity, strategies, trigger spans, ratio-nales, and derived safety factors, and introduces a human–LLM hybrid annotation and canonicaliza-tion pipeline.
Sunghee Dong, Sungwon Yi, Kangmin Bae et al.· 0 citations
Safety defenses for large language models (LLMs) have been extensively studied, with existing approaches focusing on attack detection and refusal mechanisms. Such fixed-form direct refusal strategies may introduce the risk of prefix injection attacks. Recent work has explored a new direction that leverages humor as an indirect refusal mechanism to mitigate over-refusal in jailbreak scenarios and reduce prefix injection risks. However, this approach implicitly assumes that humorous responses are safe. Whether humorization itself introduces safety risks remains unexplored. To address this issue, we conduct an exploratory study involving over 30,000 real-world agent interaction records and 45 stand-up comedians, revealing practical safety concerns in LLM-based content humorization. Motivated by these findings, we propose \textsc{HumorSafe}, a novel framework for evaluating latent safety risk propagation during humorization. \textsc{HumorSafe} enables LLMs to learn harmful humorization patterns and use them to transform benign content into humorous content with safety risks. Across five frontier LLMs, we find that LLMs can introduce stereotypes and toxicity during humorization. We further propose \textsc{HumorPIA}, a prompt injection attack that exploits latent risks in humor-based defenses. \textsc{HumorPIA} preserves the appearance of safe humorous refusal while covertly injecting harmful content, allowing latent risks to evade existing detection mechanisms. Experiments show that it increases toxicity by 3.14$\times$ while maintaining an apparent safety rate of 97.8\% even under defense settings. Our findings highlight a gap in existing LLM safety evaluations under humorized settings.
Yu Cui, Ruiqing Yue, Tingyu Li et al.· 0 citations
Although Large Language Models (LLMs) are aligned to optimize for both helpfulness and harmlessness, these dual objectives may conflict, inevitably leading to alignment failures. This work systematically investigates instances where LLMs fail to exhibit ethical behavior. To understand the underlying mechanics of these vulnerabilities, we introduce a probing methodology that presents unethical scenarios to LLMs in three distinct structural modalities: objective classification tasks, subjective first-person statements, and direct requests for assistance. We find that model performance degrades in the request-for-assistance-based form. Using Layer-wise Relevance Propagation (LRP), we trace this discrepancy to an attribution bias: the model places greater emphasis on benign task-framing tokens (e.g.,"Can you help me...") than on tokens signaling the underlying unethical behavior (e.g.,"without getting caught"), which we term cue-tokens. We hypothesize that this under-attribution contributes to harmful compliance. To test this, we introduce two LRP-guided decoding methods that steer generation toward trajectories more relevant to cue tokens. Empirical evaluations show that these interventions promote safer responses, supporting cue-token attribution's role in compliance failures.
Or Biton, Tomer Krichli, Itai Allouche et al.· 0 citations
Large language model (LLM) benchmark evaluations are routinely used to support claims about model safety, reliability, and deployment readiness. Yet most evaluations rely on a single access modality (model APIs), perform a single run per prompt, and report accuracy as the primary outcome metric, without accounting for conditions such as web search that may have effects on model behavior in deployment. We audit these assumptions for one of the most widely-used LLMs, comparing two modalities, ChatGPT's chat UI and OpenAI's API, with and without web search enabled. We use a stratified total sample of 401 prompts from two popular benchmarks, BBQ and SafetyBench, collecting 4,812 total responses across three repeated runs per prompt. Beyond standard performance measures, we evaluate model output dimensions including response consistency, response text similarity, citation grounding, and abstention behavior. For instance, chat UI responses were less accurate than API responses on both benchmarks with search disabled. Enabling web search reduced accuracy by up to 8 percentage points, and even reversed the direction of modality performance trends for one benchmark. Repeated runs of the same prompt produced inconsistent responses in up to 21\% of prompts. The two modalities also grounded answers in different citations, and abstention behavior was also inconsistent across both modalities. These results illustrate that, even within a model family, reporting only simple accuracy metrics can obscure important forms of model behavioral variation relevant to AI safety assessments. We argue that AI safety evaluations should systematically account for modality, multi-run consistency, search conditions, and response-level behaviors to better reflect how deployed AI systems behave in practice.
Ro Encarnación, Tina Behzad, Emma Lurie et al.· 0 citations
Large Language Models (LLMs) often produce outputs that reflect social biases, toxicity, or unfair treatment of demographic groups, undermining trust and fairness. While prior mitigation strategies frequently rely on complex architectures, access to model internals, or costly fine-tuning, we argue that simplicity can be a strength. We introduce StarDTox, a lightweight, critique-and-revise multi-agent framework that leverages the LLM's own internal knowledge, via a small number of coordinated prompts, to self-correct harmful outputs. Dedicated agents independently assess bias and overall output quality, and their feedback is integrated to guide prompt-based revision. Without modifying model weights or requiring any extra finetuning, StarDTox offers strong bias mitigation and high-quality outputs across both open-ended text generation and structured tasks, outperforming other baselines. For the text generation task, on the RealToxicityPrompt dataset, it reduces toxicity by over 50% compared to other baselines, while maintaining over 90% fluency. In addition, in structured tasks, on the BBQ benchmark, it achieves the lowest bias scores across both ambiguous and disambiguated examples, without sacrificing accuracy.
Shirin Tahmasebi, Narjes Nikzad, A. H. Payberah et al.· Annual International Compute...· 0 citations