Anthropic used Claude to search for ways of reducing ten kinds of model safety failure, including deception, privacy violations and harmful compliance. Instead of asking for a single proposal, its automated researcher read previous work, suggested training interventions, ran experiments and evaluated the results.
The methods improved the tested benchmarks across all ten categories without degrading the student models' measured general capabilities. Anthropic also checked some interventions against held-out evaluations that the research agent hadn't seen. In separate tests, methods developed on smaller models remained effective on models up to 4.7 times larger.
The process had restrictions. A monitoring agent reviewed each proposed method before execution, and Claude was not allowed to copy its own alignment directly into the student. In one human comparison, 28 safety researchers had up to eight hours to propose methods but couldn't iterate on their submissions, so the comparison does not establish that an autonomous agent is generally better than human researchers.