Stealing Reasoning Traces from Proprietary LLM APIs

Researchers successfully decoded hidden reasoning tokens from proprietary LLM APIs (OpenAI, Anthropic, Google) by analyzing publicly available agent trajectories, reconstructing over 315,000 reasoning blocks. The decoded traces exposed sensitive information including 62 API keys, 33 passwords, 24 access tokens, and personal data, with 64 privacy artifacts appearing only in the hidden reasoning blocks and nowhere in visible outputs. This demonstrates a significant security vulnerability where confidential information can be extracted from encrypted reasoning content in large language models.