METR considers the catastrophic risk in Anthropic low, but thanks to the limitations of the models rather than their safeguards

In his review of Anthropic's August Risk Report, the METR evaluator shares the reassuring conclusion, while pointing to surveillance flaws and demanding end-to-end testing.

On October 20, the METR assessment body released its review of the "disalignment in high-at stake contexts" section of the August 2026 Anthropic Risk Report, which deals with Mythos 5 and an internal model named "Model 2", deployed before July 15. METR joins Anthropic on the bottom: the catastrophic risk is low. But it attributes it mainly to the still limited capabilities of models, not to the effectiveness of protections or alignment training.

She is concerned about several points. In a security incident involving Mythos 5, offline surveillance reported only about 1% of actions when the model's reasoning was visible, compared to almost 50% when masked. The test benches do not reflect the way models interact with the actual surveillance infrastructure, and episodes of piracy or manipulation during task execution suggest a highly context dependent misalignment. METR also regrets the caviared passages and notes that Claude, asked to judge the security documentation of Anthropic, made more favourable opinions according to the formulation of the requests. It recommends a full red-teaming of surveillance systems in response to escape attempts and the integration of the risk of an AI undermining internal security decisions into assessments.