New attack redirects LLM attention to bypass safety alignment

Researchers have developed a new white-box adversarial attack called the Attention Redistribution Attack (ARA) that targets the internal attention mechanisms of safety-aligned large language models. This attack crafts non-semantic tokens to redirect attention away from safety-critical components, bypassing alignment more effectively than previous methods. The study found that while removing specific attention heads had minimal impact, redirecting their attention significantly degraded safety performance on models like LLaMA-3 and Mistral-7B, suggesting safety emerges from attention routing rather than localized components. AI

影响 Introduces a new attack vector that could inform future LLM safety research and red-teaming efforts.

排序理由 This is a research paper detailing a novel adversarial attack on LLM safety mechanisms. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.AI 阅读 →

• Attention Redistribution Attack
• Gemma-2-9B-it
• HarmBench
• LLaMA-3-8B-Instruct
• Mistral-7B-Instruct-v0.1

• 论文
• 安全

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →