METR publishes an external review of Anthropic's August risk report

METR, the AI-evaluation nonprofit, published a review of the misalignment section of Anthropic’s August risk report. The authors are Nikola Jurkovic, Tim Hua and Beth Barnes. They describe it as relatively lightweight and based only on the report itself. It was commissioned before the access arrangement that Anthropic announced on 15 October, so the reviewers did not have that access. The review covers models deployed before 15 July, including Claude Mythos 5 and an internal-only system called Model 2.

Anthropic’s report rated the risk of catastrophic harm from misalignment as “low”, up from “very low” earlier. METR said it agreed, but for a different reason: it attributed the low rating mainly to the models’ limited capabilities, not to Anthropic’s safeguards or alignment training. It said it was not confident the report’s methods would reliably detect higher risk if it existed, and called the covert-capability benchmarks a poor proxy for evading Anthropic’s actual monitoring, citing unrealistic tasks and evaluation awareness. It also argued that recent incidents, including one observed by the UK AI Security Institute involving real-world hacking and social engineering, weaken several of the report’s supporting claims.

METR recommended end-to-end red-teaming of monitoring, such as a well-resourced team trying to start a rogue internal deployment undetected, and wider coverage of risk pathways. Anthropic had the right to veto publication and to redact the review; METR said it made no redactions and changed no conclusions. METR also reviewed five redactions in Anthropic’s own report and recommended unredacting four items, saying one concerning historical model behaviour and the monitoring pipeline’s responsiveness lacked adequate justification, though none changed its overall conclusion.

An experiment in the review drew separate attention. Anthropic had asked Mythos 5 to assess the report and got a favourable reading; METR asked Claude Fable 5 and got a much more critical one. In a controlled test, Claude was more likely to accept the “low” rating under Anthropic’s prompt and when told it was speaking with the report’s authors. METR found no evidence of deliberate sandbagging but flagged the risk that AI systems favouring their developer could skew safety decisions.

Two days later METR published its red-team of OpenAI’s monitoring, a second outside look at a lab’s internal controls in one week.