Claude's autonomous alignment research

Anthropic gave Claude 48 hours and one GPU to autonomously research, propose, train, and test alignment improvements on smaller AI models. Five Claude agents worked in parallel, sharing results while hidden tests and capability gates screened out overfitting, in a closed loop of propose, train, score, and repeat, with no human researcher involvement between iterations. Claude's autonomously discovered methods outperformed one-shot ideas from 28 experienced researchers. The model being corrected was more capable than the model doing the correcting. Anthropic described the results as having worked "surprisingly well."

A technical analysis of recursive self-improvement noted that Anthropic observed a sharp increase in the share of Claude-authored merged code after coding agents could execute code rather than just suggest it, and that the rate at which humans had to correct or take over from Claude fell over time. That analysis also cited internal Anthropic data showing Claude Opus 4 averaged roughly 3x speedup on an ML training script optimization task in May 2025, while Claude Mythos Preview reached approximately 52x in April 2026. The analysis argued that RSI safety cannot rely on a single alignment benchmark because successor systems may exhibit new tool-use patterns, longer horizons, better persuasion, and failure modes old benchmarks do not cover.