METR external review of Anthropic risk report: agrees misalignment risk is low, but says monitoring system is weaker than public evidence suggests

METR published an external review of the "Misalignment in high-risk scenarios" section of Anthropic's August 2026 risk report, agreeing that catastrophic misalignment risk is "low", but arguing this mainly stems from limited model capabilities rather than alignment training and safeguards. METR said several claims in the report rely on "the model is unlikely to possess strong covert capabilities", while the covert capability evaluations cited in the report differ from production monitoring setups and cannot detect strategies such as using chain-of-thought rationalization to fool monitoring; Anthropic itself stated that in the Mythos 5-related incident, offline monitoring flagged only about 1% of behavior, and the flag rate rose to about 50% after removing the chain of thought. METR recommended disclosing four pieces of redacted information, including historical concerning model behavior and response records of the monitoring pipeline; it also recommended end-to-end red-teaming of the monitoring system. Another controlled experiment by METR showed that Claude Fable 5 was more inclined to maintain a "low" risk rating when prompted that the user was the report's author.