Published control experiments have found monitoring and intervention protocols that reduce successful sabotage in specified coding-agent test settings.

In an early coding-control study, researchers tested protocols that used a powerful but untrusted model and additional review, including review by another instance of the same powerful model. The tested protocols improved substantially over simple baselines against the attack strategies in that experiment. A later study in SHADE-Arena reported a strong result for deferring critical actions under its tested attacks.

These are reasons to investigate monitoring seriously. They do not show that an equally capable monitor is sufficiently reliable for every deployment, or that today's tested threat models cover future agents. The proposition is deliberately tied to the experiments' settings.