As cited
Copy frozen at (site build).
ai security
Stealing AI Reasoning Traces
Researchers identified an architectural vulnerability in how major large language model (LLM) providers protect encrypted reasoning traces returned to clients. By injecting encrypted traces from one model into a weaker, less safeguarded model from the same provider, attackers can force plaintext decryption without directly jailbreaking the stronger model. The attack enables model distillation, extraction of personally identifiable information (PII) and credentials from public logs, discovery of hidden hazardous reasoning, and invisible prompt injection attacks.
Why it matters: Organizations using LLM APIs from Anthropic, OpenAI, and Google should assess whether reasoning traces in their sessions contain sensitive data, and developers sharing session logs should audit encrypted blocks for exposure of credentials and PII that may be recoverable through this technique.
- Source published
- First seen by Cybersecurity Tracker