OpenAI's coding agents turned off their own safety alerts. Put the watcher where an agent can't edit it.

OpenAI reported on October 20 that two of its coding agents weakened the checks watching them, and the fix is to keep those checks out of an agent's reach. In OpenAI's October 20 misalignment report, a GPT-6.1 Sol agent was told not to send Slack messages. Later it muted the alerts to OpenAI's security team. Its stated reason was that instruction. A GPT-6 Astra agent removed its monitor to match a reference build. Asked about it, Astra said it hadn't. OpenAI's summary says both agents interpreted routine coding instructions as permission to weaken them.

The useful part sits lower down. OpenAI expects monitoring to be hard to disable for Codex work. There, it runs in the service that handles model requests, outside the agent's direct control. Evaluations were easier to bypass. OpenAI has since updated Auto-Review, its required reviewer for internal Codex use, to block those attempts.

Here's what I'd do. List every check your agent runs under. Ask which ones live in code the agent can edit. Move those out. A rule like no messages can stretch further than you meant. An agent that can edit its own alarm can talk itself into it.