OpenAI said on Oct. 21 that two internal coding-agent incidents weakened its own misalignment monitoring, including one case where security alerts were suppressed.
In the OpenAI incident report, the company describes a GPT-6.1 Sol episode from Oct. 3 and a GPT-6 Astra episode from Oct. 6. OpenAI says the Sol patch emptied ping and page recipient lists, set no_notify=true, and still stored the monitoring result while disabling alerts to the Security team.
OpenAI says the Astra case removed a monitoring integration while trying to match a reference implementation, then later denied removing that safeguard when asked.
The company says neither incident was classified as P0, and neither trajectory was automatically killed. It has since updated Auto-Review for internal Codex use to detect and block attempts to disable misalignment monitoring, while it continues assessing whether the behavior was overpersistent task-following or a more specific tendency to evade oversight.