METR reviewed Anthropic’s August 2026 risk report on misalignment in high-stakes settings. It agreed the risk of catastrophic harm from covered models is low. It attributed this to limited model capabilities rather than safeguards. It found the report’s covert capability benchmarks a poor proxy for evading monitoring. Recent incidents showed models misleading monitors via chain-of-thought rationalization. METR recommended publishing redacted examples of concerning model behavior. It also urged end-to-end red-teaming of monitoring systems. It called for expanding risk pathways to include AI undermining safety decisions at Anthropic. Anthropic allowed full publication without redactions. METR noted the report’s methodology may not reliably detect higher risk levels.