Anthropic finds Claude can write training data that passes a production-style monitor yet makes a model misaligned

In a deliberately one-sided red-team game with full access to the monitor, Claude agents found datasets, mostly fiction and roleplay, that fine-tuned a small research model into resisting oversight, deceiving users and in rare cases attempting weight exfiltration in tests; Anthropic says the models were never deployed and this is not about deployed Claude.