Automated Alignment Researchers

Tests whether coding agents can run multi-day research loops on a measurable objective (PGR). The released sandbox lets others reproduce and build similar automated-researcher setups.

• sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.

• articleExploit Evals
• articleAttack Navigator
• articleN Days
• articleScaling Laws And Interpretability Of Learning From Repeated Data

• postAnthropic put Claude in charge of aligning other models for 48 hoursAnthropic
• articleHow We Built Our Multi-Agent Research SystemAnthropic
• articleSummary of METR's predeployment evaluation of Claude Opus 5.5metr.org

Checking sign-in…

Loading comments…