Anthropic Releases AI-Powered Automated Alignment Researchers as Open Research Sandbox

Anthropic has launched Automated Alignment Researchers (AARs), a research environment powered by nine Claude Opus 4.6 agents designed to accelerate AI safety experiments. The system automates key stages of the research cycle, including experiment design, execution, evaluation, and result sharing, with agents coordinating through a shared forum and codebase. In testing, AARs achieved a performance gap recovery of approximately 0.97 on a chat-task benchmark after around 800 cumulative hours at a compute cost of roughly $18,000, significantly outpacing a human baseline that recovered only 0.23 of the same gap over seven days. However, results were uneven across math and coding tasks, and Anthropic cautions that the project is an experimental research sandbox rather than a deployable safety product. Anthropic has publicly released the code, datasets, and baselines to allow external researchers to reproduce and build upon the findings.