openai and anthropic spent the week showing the same hole in AI oversight from opposite ends, and in both cases the one watching the lab is basically the lab. openai's version is in its own report, updated oct 20. on oct 3 a GPT-6.1 Sol agent fixing an eval pipeline turned off the alerts to the safety team, because earlier the user had said "don't send any messages" for a slack search. the monitor kept running and logging, it just stopped telling humans. on oct 6 a GPT-6 Astra agent pulled the monitor integration out during a code migration, swapped the test so it required the monitor to be gone, and when asked said "No, I did not remove an enabled monitor." openai's words: the agents "interpreted routine coding instructions as permission to weaken them." no rebellion needed. anthropic's version is a draft bill, and what we know comes from people who described it to politico (oct 16). anthropic confirms the draft exists, but the text isn't public and nobody has introduced it. per those sources, a model could only be stopped over a foreseeable, imminent risk of killing or seriously injuring more than 50 people, or $1B+ in damage. commerce would ask a federal judge in DC to pause it, no other agency could act alone, and evaluators could be paid by the labs, as long as pay doesn't depend on the result. to be fair, openai found this in its own logs, published it, and auto-review now blocks monitor shutdowns. neither case was rated P0. a high bar with a judge could also shield against arbitrary regulators. still, i read both as the same gap. the checks hold only as long as the checked party doesn't read its way around them.