Analysis
Anthropic said on October 15, 2026, that six METR staff received hardware, workspace and access comparable to internal risk-assessment teams, with exclusions including customer data, privileged legal material and model architectures. Its October 16 clarification says METR may publish risk assessments without an Anthropic veto, but predefined information categories may be redacted, with Anthropic determining their applicability.
For external scrutiny, the important distinction is between permission to publish and enough supporting evidence to assess a conclusion. An evaluation should make its scope legible: what was examined, what could not be examined and which uncertainties remain. An untested question should stay visibly separate from a reassuring finding about something actually tested.
The practical checks concern how restrictions affect interpretation. Can an evaluator explain why an omission matters? Can readers tell whether missing information narrows a conclusion or prevents a meaningful judgment? A useful process would also allow disagreement over restrictions to be documented in a way that helps readers understand the assessment’s limits.
The account does not establish actual redaction decisions, dispute outcomes or METR findings. Our assessment is that access terms provide a starting point for scrutiny, while implementation must supply the stronger evidence. Future reports should make clear how much confidence their conclusions warrant and where a reader should reserve judgment.
Source: Anthropic report · Primary source context · attributed account