On October 15, 2026, Anthropic said six METR staff had received internal tools, hardware and workspace. METR can publish without Anthropic's veto, subject to narrowly defined redactions. Other evaluator partnerships do not automatically receive equivalent rights. Anthropic's announcement sets out the distinctions.
What independent evidence can establish
METR's October 14 assessment proposal separates three questions: whether agents would attempt harmful behavior, whether controls would catch it, and whether AI-assisted research could accelerate capability growth beyond safeguards. It explicitly says the proposal cannot provide confident assurance that development is safe.
That distinction matters for procurement. An evaluator's presence should prompt a request for evidence, not replace one. Our reading is that the valuable output is a report whose findings and limitations can be examined, rather than a partnership announcement treated as a safety seal.
Read the constraints alongside the conclusions
In its account of working with AI companies, METR commits to describing material restrictions on access, publication and evidence. It also acknowledges that voluntary cooperation leaves evaluators dependent on developers granting access, and that some information cannot be independently verified.
METR says it has declined payments and donations from frontier AI companies. Financial separation is useful context, but it does not eliminate the access problem. Readers should examine both the organization's incentives and the particular engagement's terms. METR explains both dimensions.
A useful precedent, not a blanket endorsement
SecureBio's earlier external review describes receiving an unredacted report, 110 pages of additional material and responses to more than seventy follow-up questions. It also states that Anthropic held veto and redaction rights for that engagement but exercised neither. The review's conclusions were model- and risk-specific.
The lesson for business readers is to avoid transferring assurance across models, versions or workflows. A detailed review of one risk does not answer every question about an agent connected to company systems.
What deployment teams should request next
Use external findings as one input to your own deployment decision. Ask which system was assessed, what reviewers could inspect, what remained inaccessible and whether restrictions could change the interpretation. Then test your intended workflow separately.
- Define which actions require human approval before granting tool access.
- Keep a record of model versions, permissions and incidents during the pilot.
- Assign an owner who can pause the workflow and investigate unexpected actions.
These are deployment recommendations, not controls validated by this announcement. Better visibility into a model developer does not establish that your particular integration is ready for production.