Preprint reports fewer AI agent violations with live monitoring and revocation

An arXiv preprint reports that live monitoring with permanent revocation reduced security violations from 12.00 to 3.81 per 24 assigned task opportunities against an adversarial controller that could observe the monitor's rule and statistic. The monitor tracks evidence of violations against a prespecified reference rate. The researchers also show why finite safety assessments cannot exclude attackers that preserve normal behavior until certification and intervene afterward. In simulated RoCo collaboration, that switching strategy produced commanded-separation violations in 34 of 36 post-certification episodes. The monitoring approach comes with an operational tradeoff: revocation also reduces the number of tasks executed.