Breaking Bad: Interpretability-Based Safety Audits of State-of-the-Art LLMs

• URL: http://arxiv.org/abs/2604.20945v1
• Date: Wed, 22 Apr 2026 16:51:49 GMT
• Title: Breaking Bad: Interpretability-Based Safety Audits of State-of-the-Art LLMs
• Abstract summary: We present a comprehensive, interpretability-driven jailbreaking audit of eight SOTA open-source large language models (LLMs)<n>We introduce an adaptive two-stage grid search algorithm to identify optimal activation-steering coefficients for unsafe behavioral concepts.<n>Our results establish interpretability-based steering as a powerful tool for systematic safety audits, but also highlight its dual-use risks.
• Score: 12.500730166535105
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Effective safety auditing of large language models (LLMs) demands tools that go beyond black-box probing and systematically uncover vulnerabilities rooted in model internals. We present a comprehensive, interpretability-driven jailbreaking audit of eight SOTA open-source LLMs: Llama-3.1-8B, Llama-3.3-70B-4bt, GPT-oss- 20B, GPT-oss-120B, Qwen3-0.6B, Qwen3-32B, Phi4-3.8B, and Phi4-14B. Leveraging interpretability-based approaches -- Universal Steering (US) and Representation Engineering (RepE) -- we introduce an adaptive two-stage grid search algorithm to identify optimal activation-steering coefficients for unsafe behavioral concepts. Our evaluation, conducted on a curated set of harmful queries and a standardized LLM-based judging protocol, reveals stark contrasts in model robustness. The Llama-3 models are highly vulnerable, with up to 91\% (US) and 83\% (RepE) jailbroken responses on Llama-3.3-70B-4bt, while GPT-oss-120B remains robust to attacks via both interpretability approaches. Qwen and Phi models show mixed results, with the smaller Qwen3-0.6B and Phi4-3.8B mostly exhibiting lower jailbreaking rates, while their larger counterparts are more susceptible. Our results establish interpretability-based steering as a powerful tool for systematic safety audits, but also highlight its dual-use risks and the need for better internal defenses in LLM deployment.

Related papers

• Babel: Jailbreaking Safety Attention via Obfuscation Distribution Optimized Sampling [21.801812957325662]
  Large Language Models (LLMs) remain vulnerable to jailbreak attacks.<n>We propose Babel, an efficient black-box attack framework.<n>Babel achieves state-of-the-art attack success rates and superior query efficiency.
  arXiv  Detail & Related papers  (2026-05-18T07:27:59Z)
• Persona Non Grata: Single-Method Safety Evaluation Is Incomplete for Persona-Imbued LLMs [5.478971182058342]
  Safety evaluations almost always study prompt-based personas alone.<n>We show this is incomplete: prompting and activation steering expose *different*, architecture-dependent vulnerability profiles.
  arXiv  Detail & Related papers  (2026-04-13T07:34:02Z)
• Claw-Eval: Toward Trustworthy Evaluation of Autonomous Agents [66.97968363332465]
  We introduce Claw-Eval, an end-to-end evaluation suite addressing all three gaps in agent benchmarks.<n>Claw-Eval comprises 300 human-verified tasks spanning 9 categories across three groups.<n>Every agent action is recorded through three independent evidence channels.
  arXiv  Detail & Related papers  (2026-04-07T17:43:18Z)
• Deliberative Alignment is Deep, but Uncertainty Remains: Inference time safety improvement in reasoning via attribution of unsafe behavior to base model [50.29667251847595]
  We show that despite being larger in model size and stronger in safety capability, there exists an alignment gap between teacher and student language models.<n>We propose a BoN sampling method that attributes the unsafe behavior back to the base LLMs in the latent space.<n>In particular, across 7 teacher models and 6 student models of different classes and sizes, we show an average attack success rate (ASR) reduction of 28.2% in DAN, 31.3% in WildJailbreak and 35.4 % in StrongREJECT benchmarks.
  arXiv  Detail & Related papers  (2026-04-01T02:42:41Z)
• OOD-MMSafe: Advancing MLLM Safety from Harmful Intent to Hidden Consequences [64.01706941950489]
  Current safety paradigms primarily target malicious intent or situational violations.<n>We propose shifting the safety frontier toward consequence-driven safety, a paradigm essential for the robust deployment of autonomous and embodied agents.<n>We develop the Consequence-Aware Safety Policy Optimization (CASPO) framework, which integrates the model's intrinsic reasoning as a dynamic reference for token-level self-distillation rewards.
  arXiv  Detail & Related papers  (2026-03-10T14:16:43Z)
• Provable Defense Framework for LLM Jailbreaks via Noise-Augumented Alignment [37.06683377953812]
  Large Language Models (LLMs) remain vulnerable to adaptive jailbreaks that easily bypass empirical defenses like GCG.<n>We propose a framework for certifiable robustness that shifts safety guarantees from single-pass inference to the statistical stability of an ensemble.
  arXiv  Detail & Related papers  (2026-02-02T03:26:45Z)
• What Matters For Safety Alignment? [38.86339753409445]
  This paper presents a comprehensive empirical study on the safety alignment capabilities of AI systems.<n>We systematically investigate and compare the influence of six critical intrinsic model characteristics and three external attack techniques.<n>We identify the LRMs GPT-OSS-20B, Qwen3-Next-80B-A3B-Thinking, and GPT-OSS-120B as the top-three safest models.
  arXiv  Detail & Related papers  (2026-01-07T12:31:52Z)
• Structured Uncertainty guided Clarification for LLM Agents [126.26213027785813]
  LLM agents extend large language models with tool-calling capabilities, but ambiguous user instructions often lead to incorrect invocations and task failures.<n>We introduce a principled formulation of structured uncertainty over tool-call parameters, modeling joint tool-argument clarification as a POMDP with Expected Value of Perfect Information (EVPI) objective for optimal question selection and aspect-based cost modeling to prevent redundancy.<n>Our SAGE-Agent leverages this structured uncertainty to achieve superior efficiency: increasing coverage on ambiguous tasks by 7-39% while reducing clarification questions by 1.5-2.7$times$ compared to strong prompting and uncertainty-based baselines.
  arXiv  Detail & Related papers  (2025-11-11T21:50:44Z)
• DiffuGuard: How Intrinsic Safety is Lost and Found in Diffusion Large Language Models [50.21378052667732]
  We conduct an in-depth analysis of dLLM vulnerabilities to jailbreak attacks across two distinct dimensions: intra-step and inter-step dynamics.<n>We propose DiffuGuard, a training-free defense framework that addresses vulnerabilities through a dual-stage approach.
  arXiv  Detail & Related papers  (2025-09-29T05:17:10Z)
• ORFuzz: Fuzzing the "Other Side" of LLM Safety -- Testing Over-Refusal [27.26251627767238]
  Large Language Models (LLMs) increasingly exhibit over-refusal - erroneously rejecting benign queries due to overly conservative safety measures.<n>This paper introduces the first evolutionary testing framework, ORFuzz, for the systematic detection and analysis of LLM over-refusals.
  arXiv  Detail & Related papers  (2025-08-15T05:03:26Z)
• Critique-GRPO: Advancing LLM Reasoning with Natural Language and Numerical Feedback [59.078756231841574]
  Critique-GRPO is an online RL framework that integrates both natural language and numerical feedback for effective policy optimization.<n>We show Critique-GRPO consistently outperforms supervised learning and RL-based fine-tuning methods across eight challenging mathematical, STEM, and general reasoning tasks.
  arXiv  Detail & Related papers  (2025-06-03T17:39:02Z)
• Guiding not Forcing: Enhancing the Transferability of Jailbreaking Attacks on LLMs via Removing Superfluous Constraints [81.14852921721793]
  This study aims to understand and enhance the transferability of gradient-based jailbreaking methods.<n>We introduce a novel conceptual framework to elucidate transferability and identify superfluous constraints.<n>Our method increases the overall Transfer Attack Success Rate (T-ASR) across a set of target models with varying safety levels from 18.4% to 50.3%.
  arXiv  Detail & Related papers  (2025-02-25T07:47:41Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.