Anthropic shows Claude agents can autonomously run parts of AI alignment research

Anthropic's automated alignment researchers

On Friday, Anthropic published a paper titled "Automated Researchers Can Reliably Mitigate Alignment Failures," describing an AI system that carries out a meaningful slice of alignment research on its own. According to TechCrunch, the work was led by Chen Yueh-Han, a researcher in Anthropic's fellows program, and it shows that automated systems can improve a model's behavior on a set of alignment benchmarks without hurting overall performance.

The setup, which Anthropic calls Automated Alignment Researchers (AARs), mirrors how human research typically proceeds. Per TechCrunch, each automated researcher searches the relevant literature, proposes a training method, and runs that method against a model for about 30 minutes, iterating over successive attempts. Methods that work are kept; ones that fail are discarded, letting the system operate quickly and at scale. When tested against 10 benchmarks for specific misaligned behaviors, the automated systems improved performance on every one, TechCrunch reports.

Nine agents, 800 research hours

A dev.to summary of the research adds detail on the architecture. The experiment ran nine parallel Claude Opus 4.6 instances, each equipped with a sandbox for running experiments, shared storage, a forum for coordinating with the other agents, and access to remote scoring. Together they generated hypotheses about improving the alignment of smaller models, trained and tested those models, and analyzed the outcomes.

The run accumulated roughly 800 research hours across the nine agents over five days, at an estimated cost of about $18,000 in compute and API usage — around $22 per researcher-hour, according to dev.to. TechCrunch, quoting the paper, cites a narrower figure of about $4 per hour in API inference, set against the roughly $150 per hour Anthropic says it pays human researchers. The two numbers measure different things, but both point in the same direction: automated experimentation is dramatically cheaper than the human equivalent.

What the results showed

The experiments took place in a weak-to-strong supervision setting, where a weaker model's oversight is used to improve a stronger one. Success was measured with a metric called Performance Gap Recovered (PGR). According to dev.to, the best methods recovered 0.97 of the performance gap on open-weights datasets and transferred strongly to held-out math tasks, at 0.94. Transfer to held-out coding tasks was much weaker, at 0.47, though dev.to notes this was roughly double the human baseline in some cases. A production-scale test using Claude Sonnet 4 showed only limited improvement.

The paper is direct about its own standing relative to people. TechCrunch quotes it saying the best AAR method beats what experienced humans propose on average within six hours, that human-guided research directions did not lead to stronger performance, and that the results provide early evidence automated alignment post-training could become practical in the near term.

The caveats matter

Both write-ups stress limits. TechCrunch notes the system only works to the extent that its benchmarks actually reflect alignment goals, and that establishing and maintaining those benchmarks — along with the literature the automated researchers draw from — remains significant work. Dev.to emphasizes the uneven transfer results and the weak production-scale showing, arguing the demonstration proves Claude can accelerate hypothesis generation and experimentation, not that it can define and solve alignment single-handedly. Anthropic itself says human oversight remains necessary.

Why it matters

The paper is an early, concrete step toward recursive self-improvement — AI improving the training of AI. Alignment post-training is arguably the most safety-relevant place to start, because automating it would let labs test and correct model behavior far faster than human teams can. But the results cut both ways: gains that held on benchmarks and math tasks faded on coding and at production scale, a reminder that automated improvement inherits whatever its evaluation actually measures. For now, humans still set the goals and judge the outcomes; the machines are getting faster at everything in between.