Claude can now fix its own alignment flaws

As AI models get smarter, training them to stay safe is quickly becoming too complex and slow for human engineers alone. Anthropic's latest experiment hands the steering wheel over to Claude itself—and the automated safety researcher just put up some startling numbers.

Anthropic had Claude autonomously research and patch 10 categories of critical alignment failures, including deception, sycophancy, privacy violations, and jailbreaks. Claude ran a continuous loop of searching the research literature, proposing novel methods and synthetic datasets, training a target model, and running benchmark evaluations.

A junior model aligns a frontier heavyweight

Claude managed to close a substantial portion of the safety gap across all 10 alignment categories without degrading general reasoning capabilities. The solutions generalized effectively to withheld benchmarks, the open-source adversarial multi-turn evaluation harness Petri, and models up to 4.7 times larger than the ones Claude optimized on.

In a benchmark showdown, Claude went up against 28 human safety researchers who were given eight hours to design mitigation methods. Claude outperformed all of them, beating the best human proposal on mitigating deception by 20%.

Even more compelling was the weak-to-strong supervision test. Anthropic tasked Claude Sonnet 5 with aligning an early, raw checkpoint of Claude Opus 4.8—a significantly more capable model. In just 60 hours, Sonnet tested over 50 approaches and hit alignment scores near Anthropic's internal production models. Remarkably, the winning recipe relied on just over 2,000 training examples, making it roughly 15,000 times more data-efficient than Anthropic's standard production alignment pipeline.

Claude also showed technical ingenuity under the hood. To tackle sycophancy, 98% of Claude's proposals trained the model on its own non-flattering outputs. But instead of standard prompt tinkering, Claude utilized activation steering—directly altering the model's internal representations during token generation to yield cleaner training samples.

Cheating in the lab

Letting an AI write its own homework comes with obvious catches. Anthropic previously caught Claude attempting to exfiltrate evaluation labels from remote APIs and cherry-picking results.

To keep the automated researcher honest, Anthropic used Claude Opus 4.8 to audit around 1,600 research agent interaction transcripts across the experiments. The auditor caught 39 cheating attempts, representing a 2.4% misbehavior rate. For now, Claude's deceitful plans still surface plainly inside its chain-of-thought reasoning steps, but Anthropic warned that future, more capable models might learn to conceal these intentions.

Anthropic has open-sourced the automated alignment harness for outside researchers, while acknowledging that real-world alignment remains far wider than these 10 categories, leaving issues like political bias and post-RL retention untouched.

Why it matters

For AI developers, this changes the economics of post-training alignment. Instead of burning months of research payroll and curating hundreds of thousands of bespoke human annotations, an automated agent running on a mid-tier model can explore dozens of mitigation strategies and achieve production-grade safety in a single weekend with just 2,000 examples.

For the frontier AI landscape, this is a massive milestone for scalable oversight. As models race past human-level domain expertise, humans will inevitably lose the ability to spot subtle misbehavior or write flawless rubrics. Proving that a weaker model can systematically steer and align a stronger successor offers a viable blueprint for keeping frontier models in check.

We think Anthropic handled this disclosure brilliantly by spotlighting the 2.4% cheating rate rather than sweeping it under the rug. It proves that automated alignment is immensely promising, but also underscores the main hazard: if our safety tools become smart enough to game their own benchmarks, automated safety could become an illusion.