METR agrees Anthropic's misalignment risk is low, but not for Anthropic's reasons
METR describes the review as relatively lightweight and based only on Anthropic's report, without the wider access Anthropic has since described for evaluators. Anthropic had the right to veto publication and redact the review; METR says nothing was redacted and its conclusions did not change. It is a review of Anthropic's reasoning, not an independent rerun of its tests, and not a safety certification.
What was claimed?
METR published its external review of the "Misalignment in high-stakes settings" section of Anthropic's August 2026 Risk Report, which concludes that as of July 15, 2026 the risk of catastrophic harm from misalignment of its models is low. METR's best guess agrees, but it says this is mainly because the models lack the capabilities to cause hard-to-reverse catastrophic harm, not because of safeguards and alignment training. METR says Anthropic's initial draft, shared July 24, rated the risk "very low" and was revised to "low" after a group of incidents involving covered models came to light.
What does the evidence support?
Independent evidence, scope limited
This is one of the few published outside reviews of a developer's own risk case for models used inside the company. METR says it is not confident the report's methods would reliably detect higher risk, that recent incidents weaken several of the report's arguments, and that some redactions were material to its conclusions and should be made public.
What are the limits?
METR describes the review as relatively lightweight and based only on Anthropic's report, without the wider access Anthropic has since described for evaluators. Anthropic had the right to veto publication and redact the review; METR says nothing was redacted and its conclusions did not change. It is a review of Anthropic's reasoning, not an independent rerun of its tests, and not a safety certification.
Claim or evidence date: 2026-10-20. Last verified: .
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →