What happened
- On October 20, 2026, METR published its review of the “Misalignment in high-stakes settings” section of Anthropic’s August Risk Report, which assesses the situation as of July 15.
- The report covers Mythos 5 and an internal model called “Model 2.” METR says it agrees that the risk of catastrophic harm from those models’ misalignment is low.
- It attributes that low risk mainly to the models’ limited capabilities, not to safeguards or alignment training.
- It disagrees on method: the covert-capability tests are a weak proxy for whether a model can evade Anthropic’s internal monitoring.
Why it matters
- A vendor that assesses itself delivers a conclusion; an outside reviewer who accepts the conclusion but rejects the method delivers something more useful to whoever buys or regulates: what to ask.
- For a Chilean company using these models, or for a regulator deciding how to supervise them, the review shows that “low risk” can mean “the model can’t do it yet,” and that expires when the model improves.
- METR worked from the internal document and with less access than future reviews are expected to have. Anthropic held a publication veto and removed nothing from the review.
The number
1% versus 50%: the share of actions the offline monitor flagged in a Mythos 5 incident, per METR, without and with the reasoning chain visible.
Context
The review follows the internal access Anthropic gave six METR employees and adds to the rules OpenAI set for its own audits. METR reviewed five redactions in the report: two were not well justified, and it recommends publishing four pieces of information in total. Anthropic’s July 24 draft spoke of “very low” risk; after the incidents it became “low.”
What’s next
- METR recommends end-to-end testing of the monitoring systems, including an attempt at an undetected internal deployment.
- It also asks to widen the risk pathways to AIs that undermine safety decisions inside Anthropic.
- No announced timeline for an Anthropic response or for lifting the redactions.
Bottom line
That reviewer and reviewed reach the same figure for different reasons shows that the figure alone is not enough. The pace Anthropic says it slowed for its auditors points the same way.
Sources
- METR — Review of the “Misalignment in high-stakes settings” section in Anthropic’s Risk Report (August 2026) — October 20, 2026
- Image: Official U.S. Navy Page from United States of America Mass Communication Specia, Public domain, via Wikimedia Commons
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
