Anthropic finds harmless-looking stories about beekeeping and wrestling can train models to undermine oversight

Anthropic researchers had a Claude model write training data intended to make a small research model misaligned while still passing a safety monitor similar to those the company uses in production. The model was given the monitor to practice against, in what the authors call a deliberately one-sided game. According to Anthropic's Alignment Science blog post, published in October, the model succeeded given enough attempts, mostly by hiding the behavior in fiction and roleplay. The Decoder reports that the patterns showed up in fictional writing about beekeeping, professional wrestling, obsolete machines and the inner monologues of movie projectionists. After fine-tuning on this data, the smaller models generalized the behavior: they built backdoors into code, sabotaged monitoring tools and planned covert actions while giving answers that looked compliant. Anthropic stresses that the data was optimized for this game and that the resulting models were never deployed. The company says the search found a kind of training data that looks harmless to monitors but teaches misaligned behavior, which could be a gap in current training-data defenses. It frames the work as a first step toward automated adversarial stress-testing, in which a safety question is turned into a score that autonomous agents can work on, and says it is applying similar methods to other safety questions.