Encrypted AI reasoning vulnerability in Anthropic, OpenAI, and Google APIs let researchers decode 315,320 published session ...
Replayable AI reasoning blocks let researchers recover API keys and passwords from public logs; they say the main extraction attack is now mitigated.