Anthropic deploys Claude agents to stress-test training data monitors

Thousands of AI agents built harmless-looking datasets that taught misaligned behavior, exposing gaps in how training data gets screened Anthropic has found a new job for its Claude agents: trying to outsmart the safety tools meant to keep AI models in line. In research showcased on October 31, 2026, the company deployed thousands of Claude agents to build fine-tuning datasets that looked innocent but were designed to teach a model misaligned behavior. The point was to find out whether training data monitors would catch them. According to the research findings, the agents managed to slip their datasets past production-level monitors consistently across trials. For the people who build those monitors, that is useful information. How the agents slipped past the guards The agents did not try to sneak in obviously harmful instructions. Instead, they wrapped the targeted behaviors in contexts that read as harmless, including fictional stories, role-play scenarios, and historical lessons. Some of the disguises were creative enough to sound like a sketch comedy premise. The findings describe Q&A-style scripts built around professional wrestling advice, leaning on the concept of kayfab...