On 15 October 2026, Anthropic announced internal access for six METR evaluators and publication rights without a company veto, subject to narrow exceptions for sensitive data. The announcement describes working terms; it does not present a completed assessment or a guarantee of model safety.
What is being assessed
METR's own public overview describes work on loss-of-control risks, alignment, controls over internal agents and acceleration of AI research. It distinguishes evaluations conducted with developer involvement from independent examinations of released models. That context matters: an external conversation and a systematic assessment with access to evidence are different forms of scrutiny.
Our editorial interpretation is that access and publication rights should be considered together. Evaluators might see an important weakness, yet the public cannot assess its significance from a general assurance alone. Conversely, freedom to publish becomes more useful when a report explains what information researchers could inspect and which questions remained outside their reach.
Exceptions belong in the account
According to the announcement, METR may publicly dispute whether redactions are justified without disclosing removed information. These are announced terms whose application remains to be seen.
For a reader, the useful question is what a restriction does to verifiability. In a hypothetical report, a weakness could be explained without exposing customer records or dangerous implementation details. The method, scope and grounds for the conclusion should still remain understandable. Protecting sensitive data and suppressing an inconvenient conclusion are different operations, and the distinction needs to be visible.
How to read the forthcoming work
We would separate a finding, the developer's response and a subsequent check of a claimed fix. Finding a weakness does not establish that every safeguard is useless. Announcing a fix does not establish that the same weakness has disappeared in every deployment. A meaningful assessment needs a tested version, a defined scope and an account of the questions that remain unanswered.
Organizational independence also matters alongside physical presence in an office. Readers benefit from a clear explanation of funding, access and publication rights for the particular engagement. We would not infer complete independence simply from the label external. Nor would the presence of an evaluator substitute for the evidence presented in the resulting assessment.
This is a concrete move toward deeper external scrutiny of an AI laboratory. The next evidence will be in the analyses themselves: what was tested, what limitations were disclosed and how the developer responded. SIExpo has not participated in these assessments and does not certify model safety on the strength of the announcement.
Editorially checked: 2026-10-16T07:59:41.652807+00:00 (UTC). Cover: conceptual AI illustration, not a documentary photograph.
