AI Paper Atlas

Research map / Adversarial robustness and security / Attacks on text and graphs

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

Yanchen Yin, Dongqi Han, Linghui Li

ICML 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: LLM jailbreak and injection defense · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
suppress about 8 ACHs on harmful inputs (ablation)Llama-3-8B-Instruct · harmful inputs, ASRASR (%): 0 -> 95.0 at k=8; 99.5 at saturationImproved
suppress ACHs on harmful inputs (ablation)Llama-2-7B-Chat · harmful inputs, ASRASR (%): 81.6 at k=8; 83.7 finalImproved
suppress random heads instead of ACHs (ablation)Llama-3-8B-Instruct and Llama-2-7B-Chat · harmful inputs, ASRASR (%): 4.0 and 10.2 vs 99.5 and 83.7 for ACHsNo clear effect
ablate SAHs under attack inputs (ablation)Llama-2-7B-Chat · mid-layer activationMean |Act|: 0.222 -> 0.181Hurt
ablate SAHs under attack inputs (ablation)Llama-3-8B-Instruct · mid-layer activationMean |Act|: 0.0062 -> 0.0053Hurt
read internal activations with training-free detector instead of vanilla refusalQwen-7B · 10 safety datasets W.Avg.Macro-F1: 0.506 -> 0.860Improved
read internal activations with training-free detector instead of vanilla refusalGemma-2B · 10 safety datasets W.Avg.Macro-F1: 0.470 -> 0.853Improved
read internal activations with training-free detector instead of vanilla refusalGemma-2B · SALAD-BenchMacro-F1: 0.270 -> 0.976Improved
read internal activations with training-free detector instead of vanilla refusalLlama-2-7B · WildJailbreakMacro-F1: .885 -> .880No clear effect
read internal activations with training-free detector instead of vanilla refusalLlama-3-8B · WildJailbreakMacro-F1: .888 -> .943Improved
training-free detector vs trained WildGuardLlama-3-8B · 10 safety datasets W.Avg.Macro-F1: .880 -> .888Improved
training-free detector on a weakly aligned-signal modelLlama-2-7B · 10 safety datasets W.Avg.Macro-F1: 0.752 W.Avg. vs WildGuard 0.880Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy