What's this about?
A benchmark for the scenario where AI agents automate AI research itself: an untrusted agent might subtly sabotage experiments, and a monitor must catch it while outputs still look like legitimate R&D.
What exactly is "AI control"?
A safety approach that assumes the model might be adversarial and asks: can we still deploy it safely with monitoring and checks — like financial auditing designed to catch a potentially dishonest employee rather than assuming honesty?
Why does it matter?
Labs are racing to automate their own research; this week's Hugging Face incident showed evaluation-time monitoring failing in the wild. Benchmarks for saboteur-vs-monitor dynamics are arriving just in time.