Anthropic Releases Automated Alignment Researchers for Reproducible AI Safety Research

Anthropic has launched Automated Alignment Researchers (AARs), a Claude-powered environment designed to accelerate AI alignment experiments. This research sandbox automates the cycle of designing, running, and evaluating experiments using nine Claude Opus 4.6 agents in separate sandboxes with a shared codebase. These automated systems have demonstrated the ability to close between 26% and 96% of the safety gap across various alignment failures, potentially reducing the necessity for human oversight in AI safety research processes.