Replayable AI reasoning blocks let researchers recover API keys and passwords from public logs; they say the main extraction attack is now mitigated.