Review of the "Risks from automated R&D" section in the Anthropic Risk Report (February 2026)

An external reviewer can agree with a risk conclusion while finding its supporting argument inadequate. In METR's review summary, the organization criticizes how Anthropic used an internal survey to assess risks from automated research, including counting a missing answer as a negative response.

METR also argues that substantial research acceleration could matter before full automation. It nevertheless agrees that the specified catastrophic risk from Opus 4.6 or a less capable model is very low, drawing on additional evidence beyond the original report. The useful distinction is between confidence in this conclusion and whether the report's method supplies a convincing basis for it.