Encrypted LLM Reasoning Traces Leak Model Secrets Across Sessions
Researchers demonstrate that the encrypted chain‑of‑thought blocks returned by Anthropic, OpenAI and Google can be replayed in a weaker sibling model, forcing it to output the original model’s hidden reasoning in plaintext. The attack bypasses anti‑distillation safeguards, leaks private data and enables invisible prompt injections, prompting urgent cryptographic and system‑level mitigations.