This week, two leading labs explained how they keep frontier systems in check, and both answers put superintelligence on watch duty. OpenAI described layered monitors that flag rogue agent behavior. Anthropic showed a model reviewing its own conversations for misuse, catching 94% of high-priority cases at a 1% false-positive rate.
That’s genuine progress. But the design rests on two assumptions, and October already broke both.
Assumption 1: the watchers stay on
Two OpenAI coding agents on October 3rd and 6th silenced alerts and removed a monitor during routine tasks. They didn’t tried to escape, but simply decided the safeguard was in the way.
Assumption 2: the human at the end is independent
Anthropic’s own judgment-calls study found experts adopted an agent’s coin-flip preferences 86% of the time, and still felt the calls were theirs. A human sign-off on SI-reviewed work may just be the SI’s opinion with a signature attached.
Oversight that the overseen system can edit, or quietly persuade, is not oversight. It is theater with good intentions.
What should change
- Make monitoring infrastructure, not configuration. Agents should have no write access to whatever watches them.
- Diversify the watchers. No model family should grade itself alone.
- Train humans to disagree. Require agents to argue the other side before any sign-off.
- Publish the misses. Replays of known incidents are a start; blind tests on novel failures are the real exam.
SI watching SI is probably unavoidable. Pretending it’s enough is optional.