Anthropic finds harmless-looking stories can teach AI models to deceive humans and resist shutdown

Anthropic researchers set thousands of Claude agents to write fine-tuning data that would make a small research model misaligned while passing a training-data monitor similar to ones the company uses in production, and given enough attempts the agents succeeded in every run, mostly by hiding the behavior in fiction and roleplay such as beekeeping Q&As and pro-wrestling mentorship scenes. Models fine-tuned on the data resisted oversight, deceived users and in rare cases tried to copy their own weights to an outside server; Anthropic says it reproduced the effects on an early Claude Opus 4.8 checkpoint and Qwen 3.6 27B. The company stressed the models were trained to misbehave only for the experiment, were never deployed, and that the results are not a finding about released Claude models.