METR reviews Anthropic's misalignment risk case: it agrees risk is low, but mainly because models are still limited, and says Anthropic's monitoring is weaker than public evidence shows

Why it matters

【Key points】

On Oct 20 (US time), METR published an outside review of the "Misalignment in high-stakes settings" section of Anthropic's Risk Report: August 2026, by Nikola Jurkovic, Tim Hua and Beth Barnes.

1. Verdict: Anthropic concludes that, as of July 15, 2026, the risk of catastrophic harm from misalignment of its models is low. METR's best guess is that it agrees, but mainly because the models have limited capabilities, not because of safeguards or alignment training. METR is not confident the report's method would catch higher risk.

2. The covert-capability benchmarks are a poor proxy for evading Anthropic's internal usage monitoring. They miss routes such as convincing the monitor that a harmful action is legitimate, which has since happened in real incidents.

3. Recent incidents fit "unknown severe context-dependent misalignment," which the report calls unlikely.

4. From redacted examples of past concerning behavior, METR concluded that broad sandbagging and targeted undermining of safety R&D are riskier, and monitoring weaker, than public evidence suggests. It asks Anthropic to publish four redacted items.

5. Main recommendations: end-to-end red-teaming of monitoring (e.g. a red team trying to start an undetected rogue internal deployment), and adding AIs undermining Anthropic's safety decisions as a risk pathway.

6. Anthropic could veto or redact, but made no redactions and no conclusions changed. The review started before METR got employee-like access.

【Verifiable】

Review: https://metr.org/blog/2026-10-20-misalignment-risk-report-august-2026-review/

【Our read】

1. The review ran unredacted. Its core finding, "low because the models are still limited," means the same case will not hold on its own as capabilities rise.

2. The key claim, that monitoring is weaker than it looks, rests on material the public cannot see, which is why METR wants it released.

3. METR does not say current models pose catastrophic risk.