OpenAI, Anthropic, Google API Flaw Let Weaker AI Models Decode Stronger Models' Reasoning
Summary: A newly disclosed flaw in how OpenAI, Anthropic, and Google carried encrypted reasoning between API calls let researchers replay opaque reasoning blocks across sessions, users, and models, including handing them to weaker models in the same family to transcribe the hidden content. Across 6,708 public agent trajectories the team decoded 315,320 thinking blocks and counted 704 distinct privacy artifacts from genuine user sessions, including 62 API keys, 33 passwords, 24 access tokens, and seven private keys.
What you should know: Of those 704 artifacts, 64 appeared only inside hidden reasoning and nowhere in the visible trace, so sanitizing the readable conversation would still have left secrets in an opaque block another account could replay. The researchers say the attacks stopped working after mitigations, but no public acknowledgment from the three providers has surfaced, leaving that claim resting on the team's own reproducibility statement.
Read the original at The Hacker News ↗
- Topics
- AI Security
- Related
- SECURITY AFFAIRS AI-CYBERSECURITY NEWSLETTER ROUND 8 Cyber Risks · Nov 15, 2026Pentagon Renames CDAO ‘Super Intelligence Office’ Cyber Risks · Nov 15, 2026186 Fake AI Domains Redirect Users to Click Fix Lures, Browser Malware, and Search Hijackers Cyber Risks · Nov 12, 2026