METR has reviewed the misalignment section of Anthropic's August 2026 Risk Report. The October 20 review makes a distinction that matters for anyone reading AI safety claims: agreeing with a current risk assessment is different from trusting the method to detect substantially greater risk.
What METR agreed with
METR's best guess agrees that catastrophic misalignment risk was low for the models covered by the report. It attributes that judgement primarily to limited model capabilities rather than confidence in safeguards and alignment training.
That conclusion is scoped to the models and evidence under review. It should not be converted into a general assurance about future systems, nor into a claim that METR concluded a catastrophe was imminent.
Where the review finds uncertainty
The reviewers question whether the report's methods would reliably identify a higher-risk situation. They argue that covert-capability benchmarks are an imperfect proxy for evading an actual internal monitoring system. A benchmark can omit routes to failure that appear when components operate together.
METR describes this as a lightweight review based on the internal risk report. Its publication explains the access limitations and says Anthropic could veto or redact publication. METR reports that no redactions were made and its conclusions were not changed during that process.
What a stronger assessment would test
The recommendations include testing the monitoring system from end to end and broadening the pathways considered in safety-relevant decisions. For readers, that suggests asking what was tested together, which behaviours were outside the test, and how the system responds after a warning.
The review is useful because it keeps a conclusion and its confidence separate. An evaluation can contain thoughtful uncertainty and still leave important open questions. Understanding those gaps is more informative than reducing the document to a single safe-or-unsafe label.