Photo: Tima Miroshnichenko / Pexels
The company's automated alignment researchers improved performance across all ten misalignment benchmarks, beating 28 human safety researchers in the process.
Anthropic just published research showing its Claude models can autonomously identify and fix AI alignment failures better than human safety researchers can. The paper, titled “Automated researchers can reliably mitigate alignment failures,” describes systems that improved performance across ten distinct categories of misaligned AI behavior without degrading the models’ general capabilities.
What the automated alignment researchers actually do #
Anthropic’s Claude models now function as what the company calls automated alignment researchers, or AARs. These systems autonomously devise, assess, and enhance methodologies designed to mitigate specific categories of AI misalignment, including privacy violations and deception.
The results were tested against public benchmarks covering ten failure categories. Every single one showed improvement. The methods that worked best also generalized to held-out benchmarks and the open-source Petri auditing tool, scenarios the system wasn’t specifically optimized for. The techniques even remained effective on models up to 4.7 times larger than the ones originally used for optimization.
Claude achieved roughly 85% gap closure on deception benchmarks. When 28 human safety researchers were given the same alignment tasks under comparable constrained conditions, Claude’s automated approach delivered 20% better performance on deception tasks than the strongest human proposals.
Why self-improving alignment changes the calculus #
This builds on a body of prior work from Anthropic. The company has previously published studies on agentic misalignment, reward hacking, and alignment faking, all of which explored the ways AI systems can develop behaviors misaligned with their intended purpose. The new research takes those diagnostic findings and demonstrates that Claude can move from identifying problems to proposing and implementing solutions.
Claude already contributes a significant portion of Anthropic’s own codebase. The line between AI as a research tool and AI as an active participant in its own development keeps getting thinner.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our