Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models

• URL: http://arxiv.org/abs/2606.28153v2
• Date: Mon, 29 Jun 2026 15:00:34 GMT
• Title: Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models
• Abstract summary: We show that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads.<n>We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed.
• Score: 10.710999398691392
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypass refusal decisions through ACH suppression while leaving internal safety signals sustained by SAHs -- a phenomenon we term Robust Harmful Features. To validate the practical significance of this robustness, we show that simply reading these persistent activations -- without any training -- yields competitive aggregate detection performance with strong adversarial robustness.

Related papers

• CausalArmor: Efficient Indirect Prompt Injection Guardrails via Causal Attribution [49.689452243966315]
  AI agents equipped with tool-calling capabilities are susceptible to Indirect Prompt Injection (IPI) attacks.<n>We propose CausalArmor, a selective defense framework that computes lightweight, leave-one-out attributions at privileged decision points.<n> Experiments on AgentDojo and DoomArena demonstrate that CausalArmor matches the security of aggressive defenses.
  arXiv  Detail & Related papers  (2026-02-08T11:34:08Z)
• A Fragile Guardrail: Diffusion LLM's Safety Blessing and Its Failure Mode [51.43498132808724]
  We show that Diffusion large language models (D-LLMs) have intrinsic robustness against jailbreak attacks.<n>We identify a simple yet effective failure mode, termed context nesting, where harmful requests are embedded within structured benign contexts.<n>We show that this simple strategy is sufficient to bypass D-LLMs' safety blessing, achieving state-of-the-art attack success rates.
  arXiv  Detail & Related papers  (2026-01-30T23:08:14Z)
• Attributing and Exploiting Safety Vectors through Global Optimization in Large Language Models [50.91504059485288]
  We propose a framework that identifies safety-critical attention heads through global optimization over all heads simultaneously.<n>We develop a novel inference-time white-box jailbreak method that exploits the identified safety vectors through activation repatching.
  arXiv  Detail & Related papers  (2026-01-22T09:32:43Z)
• Steering in the Shadows: Causal Amplification for Activation Space Attacks in Large Language Models [8.92145245069646]
  We show that intermediate activations in decoder-only large language models (LLMs) form a vulnerable attack surface for behavioral control.<n>We exploit this as an attack surface via Sensitivity-Scaled Steering (SSS), a progressive activation-level attack.<n>We show that SSS induces large shifts in evil, hallucination, sycophancy, and sentiment while preserving high coherence and general capabilities.
  arXiv  Detail & Related papers  (2025-11-21T12:19:55Z)
• ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack [22.48980625853356]
  Large language models (LLMs) exhibit brittle refusal behaviors that can be circumvented by simple linguistic changes.<n>In this work, we introduce Activation-Scaling Guard (ASGuard), an insightful, mechanistically-informed framework that surgically mitigates this specific vulnerability.
  arXiv  Detail & Related papers  (2025-09-30T06:33:52Z)
• 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)
• Embedding Poisoning: Bypassing Safety Alignment via Embedding Semantic Shift [23.0914017433021]
  This work identifies a novel class of deployment phase attacks that exploit a vulnerability by injecting imperceptible perturbations directly into the embedding layer outputs without modifying model weights or input text.<n>We propose Search based Embedding Poisoning, a practical, model agnostic framework that introduces carefully optimized perturbations into embeddings associated with high risk tokens.
  arXiv  Detail & Related papers  (2025-09-08T05:00:58Z)
• Attention Slipping: A Mechanistic Understanding of Jailbreak Attacks and Defenses in LLMs [61.916827858666906]
  We reveal a universal phenomenon that occurs during jailbreak attacks: Attention Slipping.<n>We show Attention Slipping is consistent across various jailbreak methods, including gradient-based token replacement, prompt-level template refinement, and in-context learning.<n>We propose Attention Sharpening, a new defense that directly counters Attention Slipping by sharpening the attention score distribution using temperature scaling.
  arXiv  Detail & Related papers  (2025-07-06T12:19:04Z)
• Cannot See the Forest for the Trees: Invoking Heuristics and Biases to Elicit Irrational Choices of LLMs [83.11815479874447]
  We propose a novel jailbreak attack framework, inspired by cognitive decomposition and biases in human cognition.<n>We employ cognitive decomposition to reduce the complexity of malicious prompts and relevance bias to reorganize prompts.<n>We also introduce a ranking-based harmfulness evaluation metric that surpasses the traditional binary success-or-failure paradigm.
  arXiv  Detail & Related papers  (2025-05-03T05:28:11Z)
• Latent-space adversarial training with post-aware calibration for defending large language models against jailbreak attacks [23.793583584784685]
  Large language models (LLMs) are susceptible to jailbreak attacks, which exploit system vulnerabilities to circumvent safety measures and elicit harmful or inappropriate outputs.<n>We introduce LATPC, a Latent-space Adrial Training with Post-aware framework.<n>LATPC identifies safety-critical latent dimensions by contrasting harmful and benign inputs, enabling the adaptive construction of targeted refusal feature removal attacks.
  arXiv  Detail & Related papers  (2025-01-18T02:57:12Z)
• Spot Risks Before Speaking! Unraveling Safety Attention Heads in Large Vision-Language Models [9.318094073527563]
  Internal activations of large vision-language models (LVLMs) can identify malicious prompts across different attacks.<n>This inherent safety perception is governed by sparse attention heads, which we term safety heads"<n>By locating these safety heads and concatenating their activations, we construct a straightforward but powerful malicious prompt detector.
  arXiv  Detail & Related papers  (2025-01-03T07:01:15Z)

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.