Reilly Haskins

We run lots of evals at METR. Sometimes, agents attempt harmful actions.

I built a monitor that blocks suspicious tool calls until a human reviews them.

Writing out a case for why it's effective surfaced hidden assumptions. I'd recommend it to anyone building monitors!

Could future models learn that their CoT is being monitored and hide their reasoning to evade detection?

In our new paper, @JoshAEngels, @bilalchughtai_, and I find that yes, models finetuned on docs describing a CoT monitor evade detection far more often than unaware models 🧵

1/ We trained models on docs describing a CoT monitoring setup. Without explicit instruction to obfuscate, gpt-oss-120b and kimi k2 learned to rewrite their CoT to drop detection rates by 42-68%, while still performing a deception task..