METR’s October 20, 2026, review of Anthropic’s August Risk Report was based solely on the report, not on the employee-like access Anthropic announced five days earlier. METR’s best guess was that catastrophic misalignment risk was low for the models covered, largely because of their limited capabilities. It also cautioned that the report’s methods might not reliably detect a higher level of risk.
What METR’s review examined
METR reviewed the “Misalignment in high-stakes settings” section of Anthropic’s August 2026 Risk Report. The assessment covered models Anthropic had deployed internally or externally before July 15, 2026, including Mythos 5 and the internal-only model called “Model 2.” METR said its review relied solely on the report and had been arranged before Anthropic announced the new access on October 15.
That timing matters: METR’s October 20 conclusions are an assessment of Anthropic’s report, not findings from the newly embedded team’s access to company systems.
What Anthropic said METR could access
Anthropic said on October 15 that it had begun providing six METR employees with company hardware, office space, and access to internal tools and systems, roughly comparable to the access its own risk-assessment teams receive. Anthropic also said the small embedded team works under company security standards and enforceable agreements, and is kept separate from other METR employees.
Anthropic described three evaluator tiers, with different publication rules:
| Tier | Publication terms described by Anthropic |
| 1 | Anthropic approval is required before publication. |
| 2 | Evaluators may publish without approval, subject to specified redactions. |
| 3, METR’s arrangement | Publication rights are similar to tier 2. Anthropic says it cannot veto publication, subject to specified redactions. |
The tier labels distinguish publication rights from access. Anthropic described METR’s tier 3 arrangement as combining employee-like access with publication terms similar to tier 2.
Anthropic listed customer information, attorney-client privileged information, and model-architecture information among the redaction categories. It says it ultimately decides whether information falls within those categories. METR may make certain public statements about the access and information it received. In practice, the arrangement gives METR publication rights without an Anthropic veto, while leaving Anthropic the final decision on whether a redaction category applies.
For background on the earlier proposal for external evaluators, NeoTeo previously covered Dario Amodei’s call to slow frontier AI development.
What METR concluded, and what it wants tested
METR’s qualified low-risk judgment applied to the models covered by Anthropic’s report. The review attributed that judgment mainly to the models’ limited capabilities, rather than to safeguards or alignment training. METR also said it was not confident the report’s methods would reliably identify a higher risk level in advance.
The review recommended more comprehensive, end-to-end red-teaming of Anthropic’s asynchronous monitoring system. Red-teaming means deliberately testing a system for weaknesses. METR said the exercises should cover more of the strategies a model might attempt, and recommended including the possibility that AI systems could undermine safety-related decisions at Anthropic in the risk analysis.
