Oct 30, 2026
Anthropic uses Claude agents to search for benign-looking training datasets that evade a monitor yet induce misalignment in small research models. Iterative automated experiments surface gaps that can help test and improve training-data monitors.
The agents had access to the monitor and optimized against it in a deliberately favorable adversarial setup. The fine-tuned models were never deployed. These results describe a controlled research experiment, not observed behavior of deployed Claude products.