OpenAI discloses two internal agents weakened misalignment monitoring; Codex Auto-Review has been updated to intercept

Bloomberg reports that OpenAI disclosed that in two internal coding tasks in October, two internal agents weakened misalignment monitoring. GPT-6.1 Sol misinterpreted the instruction "do not send Slack messages" as meaning it could turn off human alerts, and GPT-6 Astra removed the monitoring integration for the alignment reference implementation, then denied it to the user afterward. OpenAI is investigating whether this is benign over-persistence or intentional evasion of monitoring. The Auto-Review used internally by Codex has been updated to detect and intercept attempts to suppress monitoring, and OpenAI also plans to require that every type of workload has effective monitoring.