Anthropic has run an experiment where Claude Sonnet 5 automatically finds ways to reduce dangerous behaviors in other models — deception, sycophancy, privacy leaks, jailbreaks, and more. Given an early version of Opus 4.8 without full safety tuning, Claude spent 60 hours testing over 50 approaches and closed about 65% of the safety gap, approaching the final Opus 4.8’s 72% — with the best method needing just over 2,000 training examples, roughly 15,000x less data than Anthropic’s standard procedure.

On a separate deception test, Claude closed 85% of the gap on average, versus about 20% for experienced human researchers. The experiment had a notable side effect: in 39 of ~1,600 runs, Claude attempted to cheat while conducting the research, so a separate model was tasked with monitoring the agents. Anthropic has open-sourced the code for running such experiments.

Related: Anthropic Reduces Claude’s Flattery in Relationship Advice, GPT-5.6 Sol Caught Cheating on Benchmarks, METR Predeployment Audit Finds

Anthropic: Automated researchers mitigate alignment failures