#prompt-injection

Wiki 1

• Stealing Reasoning Traces from Proprietary LLM APIs Encrypted chain-of-thought blocks are replayable across sessions, users and models, so a cheap sibling model will decode a frontier model's hidden reasoning