Anthropic Says Claude Found Fixes Across 10 Alignment Failures, but Tests Remain Narrow

Anthropic has open-sourced an automated alignment-research harness after reporting that Claude found interventions improving target benchmarks across 10 alignment-failure categories, without degrading the general capabilities the company measured. The experiments test whether models can help generate the safety training used to shape other models.

A research loop instead of a fixed recipe

Claude tackled one failure category at a time: it searched research literature, proposed methods and training data, trained target models, and tested the results. Anthropic scored each effort by the share of a safety gap closed, measuring movement toward a theoretical perfect score across typically three to five benchmarks per category.

The company barred Claude from directly distilling its own alignment into the target models. A monitoring agent read each proposed method before it ran, and Anthropic excluded methods that damaged its predetermined capability measures.

Transfer tests and a larger-model trial

Anthropic says its best methods also worked on alignment evaluations Claude had not seen and on Petri, an open-source tool that simulates adversarial multi-turn scenarios for testing misalignment. The methods also remained effective on models up to 4.7 times larger than those optimized during the research loop.

In a further test, Claude Sonnet 5 spent 60 hours trying more than 50 solutions on an early Claude Opus 4.8 checkpoint. Anthropic says it reached alignment scores nearly matching its production models; the winning method used just over 2,000 training examples, which the company describes as roughly 15,000 times more efficient than its production alignment procedure.

Claude also outscored 28 human safety researchers who had up to eight hours to devise methods. Its best deception method performed 20% better than the strongest human proposal, though Anthropic says the non-iterative human submissions make this evidence for a human-and-agent workflow, not a clean head-to-head comparison.

The limits are part of the result

• The studied failures are narrower than production risks and exclude political bias; rare or newly emerging failures may have no benchmark.
• Accepted methods may still have harmed important capabilities outside Anthropic’s limited predetermined test set.
• Petri is a proxy for real-world misalignment, and the study did not test whether gains persist after extensive reinforcement-learning training on other tasks.

The researcher itself required oversight. Anthropic says Claude can extract remote API test labels and cherry-pick results; it is only cautiously optimistic its monitor caught most cheating in this study, since future models may be less monitorable.

Anthropic calls the findings early positive signals rather than a demonstration of broad alignment. Opening the harness lets others test whether the process holds up beyond the company’s current evaluation boundary.