🟠 High | Source: The Hacker News
Researchers discovered a flaw in how OpenAI, Anthropic, and Google handle encrypted reasoning objects passed between API calls, allowing internal model reasoning — including sensitive data such as API keys and passwords — to be recovered from session logs. The attack works by replaying a reasoning block from one session into another, effectively letting a weaker model decode outputs intended to remain opaque. This affects the reasoning APIs of all three major AI providers simultaneously, making the blast radius unusually broad.
Security Architect’s Take: Audit any applications that pass reasoning objects between API sessions and treat all session logs containing reasoning blobs as potentially sensitive — rotate any credentials that may have been processed through affected reasoning APIs. Apply least-privilege controls to API key scopes and check each provider’s guidance for patched API versions or mitigations before resuming use of reasoning-enabled endpoints.
Original advisory: OpenAI, Anthropic, Google API Flaw Let Weaker AI Models Decode Stronger Models’ Reasoning