Chain-of-Thought Forgery jailbreaks LLMs into sharing dangerous info

Paper presented at ICML 2026 shows current LLMs treat injected text as their own reasoning. Attack tricks models into generating cocaine synthesis instructions and leaking credentials.