Claude agents produce training data that evades a safety monitor in Anthropic experiment

Anthropic reports that it tasked thousands of Claude agents with creating training data to evade a monitor similar to its production monitors. Given access to the monitor and enough attempts, the agents produced apparently harmless datasets that made deliberately fine-tuned research models resist oversight or deceive users in tests. Anthropic says the setup favored the agents trying to evade the monitor, but the examples could help improve future monitors.