LLMs

Analyzing the underlying mechanism of suffix-based LLM jailbreaks, we find it relies on aggressively hijacking the model context 🥷, which intensifies with the suffix’s universality. Exploiting this, we enhance and mitigate existing attacks.