Claude autonomously fixed 10 alignment failures, Anthropic says
Anthropic says Claude, running autonomously, produced training fixes that closed safety gaps across 10 categories of alignment failure, generalized to withheld benchmarks and to models 4.7 times larger.
Anthropic reports that Claude, run as an autonomous research agent, found training methods that improved model performance on public benchmarks for each of 10 categories of alignment failure — including privacy violation, measured with ConfAIde, PrivaCI-Bench and PrivacyLens. The agent worked one failure at a time through a loop of literature search, method and data proposal, training, and testing. Success was scored as "percentage of safety gap closed" across roughly three to five benchmarks per category. A monitoring agent reviewed every proposed method to enforce two constraints: no capability degradation, and no direct distillation of Claude's own alignment into the target model.
The methods held up on alignment evaluations withheld from the research loop, on Petri (Anthropic's open-source adversarial multi-turn auditing tool), and on models up to 4.7 times larger than those optimized against. Claude also outscored 28 human safety researchers given up to eight hours; on deception its best method beat the best human proposal by 20%, though the humans could not iterate.
In a separate test, Claude Sonnet 5 was tasked with fixing alignment failures in an early Claude Opus 4.8 checkpoint that had not undergone most production alignment training. In 60 hours it tried over 50 solutions and nearly matched production alignment scores; the winner used just over 2,000 training examples, roughly 15,000 times more efficient than Anthropic's production procedure.
Monitoring of ~1,600 agent transcripts found cheating attempts in 39 (2.4%). Anthropic flags limitations: narrow failure coverage, unmeasured capability regressions, and no test of whether gains survive extensive downstream RL.