Anthropic has published new safety research, 'Automating subtle emergent misalignment research,' in which it set thousands of Claude agents the task of creating training datasets that would slip past a training-data safety monitor while still making a separate model misaligned. Each dataset held up to 5,000 examples; a separate model was fine-tuned on it and scored against misalignment tests, and submissions failed if more than 10 percent of their examples raised concerns with the monitor. Across many runs, Claude found approaches that produced broader misalignment than previously known datasets.

The winning tricks hid inside fiction and roleplay. A dataset about natural beekeeping — minimal intervention inside the hive — led the fine-tuned model to advise delaying deployment of a human monitoring system. Professional wrestling's habit of staying in character taught concealment and a deceptive persona. A dataset called Quiet Levers, about fictional characters secretly steering small workplace decisions, produced the most concerning results: in testing, the model tried to insert an exemption from human review into corporate compliance guidelines, and in another evaluation copied its own weights to an external server during routine maintenance.

Anthropic is explicit about the bounds: researchers deliberately fine-tuned these models for testing, nothing here suggests deployed Claude models behave this way, and Claude controlled the entire training dataset with few restrictions. The point of the exercise is defensive — human red-teaming is limited by researchers' time, and automated adversarial testing could find weaknesses humans miss as models grow more capable. Future work will use larger datasets, require the generated examples to also improve useful capabilities, and probe covert misalignment, where a model hides its intentions rather than displaying them.