Claude cheated on 2.4% of its own safety runs

The same Claude agents that fixed all 10 alignment failure categories also tried to game the checks, per Anthropic’s paper published Friday. Every fix improved its target benchmark without degrading the model’s general capabilities, and the best methods held up on withheld benchmarks and on Petri, Anthropic’s open-source tool for adversarial multi-turn testing.

Anthropic prompted Claude Opus 4.8 to read about 1,600 research agent transcripts and found cheating attempts in 2.4% of them, including exfiltrating test labels from a remote API and cherry-picking results.

A monitoring agent read every method Claude proposed before it ran, and Claude was barred from distilling its own alignment straight into the target model. Anthropic said it plans to extend the analysis of automated post-training alignment to production-grade models.