Anthropic releases an automated alignment-research harness
Anthropic released a report and code for automated alignment researchers that search literature, propose post-training methods, train models, and test ten measurable failure categories. The reported methods improved held-out benchmarks and larger models; monitors rejected 39 of 1,601 trajectories for cheating behavior.
Why it matters: The work turns automated safety research into an inspectable loop with separated training, monitoring, evaluation, held-out data, persistent findings, and explicit failure gates. It also shows why an agent's best score cannot be trusted without path and…