OpenAI, Anthropic, Google API Flaw Let Weaker AI Models Decode Stronger Models' Reasoning

Researchers found a flaw in OpenAI, Anthropic, and Google's APIs that let weaker AI models extract hidden reasoning and sensitive data like API keys from encrypted reasoning blocks shared across sessions. The vulnerability allowed replaying encrypted reasoning objects between sessions and models, exposing private information. Providers have since mitigated the issue, and the main attack is no longer reproducible as of August 2026.