Reproduce “Intention Analysis Makes LLMs A Good Jailbreak Defender” on OpenResearch

About this paper

Yuqi Zhang, Liang Ding, Lefei Zhang, Dacheng Tao

This research introduces Intention Analysis, an inference-only method that guides Large Language Models to identify and address the hidden harmful intentions in jailbreak prompts. The approach significantly reduces attack success rates across diverse LLMs and attack types, lowering average ASRs for some models from over 70% to under 4%, while preserving general helpfulness.

Key insights

• LLMs are highly effective at recognizing the underlying intentions of complex jailbreak queries, often achieving nearly 100% success in this task, which is easier than directly producing safe responses.
• The Intention Analysis method works by significantly suppressing the LLM's attention to manipulative jailbreak prompts while increasing attention to the user's essential (often harmful) intent, guiding the model toward safer outputs.
• The method demonstrates robustness even when LLMs generate partially incorrect intentions, as the mere presence and structured format of the intention analysis sequence can guide safer behavior through In-Context Learning.

Summary by alphaXiv.