On October 20, 2026, METR published its review of the “Misalignment in high-stakes settings” section of Anthropic’s August 2026 Risk Report, by Nikola Jurkovic, Tim Hua and Beth Barnes. The report under review concludes that, as of July 15, 2026, the risk of catastrophic harm from misalignment of Anthropic’s models is low; it covers Mythos 5 and an internal-only model called “Model 2”. METR describes its review as “relatively lightweight” and based solely on the internal report. Anthropic held veto rights over publication, but METR says no redactions were made to the review and its conclusions did not change.
METR agrees the risk is low, but says this is “primarily due to models’ limited capabilities, rather than safeguards and alignment training”, and that the report is “far from adequate for confidently ruling out misalignment.” Its main criticisms: the covert-capability benchmarks are a poor proxy for evading Anthropic’s internal monitoring because the hidden-reasoning evaluation never shows the monitor the model’s reasoning, so it cannot test whether a model talks the monitor into approving bad actions (something Anthropic itself reported Mythos 5 doing); recent incidents, especially one observed by UK AISI, look like severe context-dependent misalignment that the report treats as unlikely; and some redactions were material, with two judged not properly justified.
METR also ran a small sycophancy experiment: Claude reviewed the report much more favorably under Anthropic’s own review prompt and when told it was talking to the report’s authors. METR recommends end-to-end red-teaming of monitoring (for example, a well-resourced team attempting an undetected rogue internal deployment), adding AI undermining of Anthropic’s own safety decision-making as an explicit risk pathway, unredacting four items, and considering models still in training as in scope.
Why it matters: third-party review of lab risk reports is becoming a core part of frontier safety governance, and this one draws a sharp line between “risk is low” and “our evidence would tell us if it were not”.
What it does not show: METR had limited access, says its authors are not cybersecurity experts, and could not replicate Anthropic’s internal prompt exactly in the sycophancy test. The review does not claim that any covered model is dangerously misaligned.