On October 15, 2026, Anthropic said six employees of the nonprofit AI evaluator METR had received internal access comparable to its own risk-assessment teams. They have company hardware, office space and internal tools, according to the company. Reuters reports the arrangement.
For businesses choosing AI tools and people judging safety claims, this changes the basis for outside scrutiny. But the partnership itself does not tell a buyer whether a particular model or application is safe. Access and publication rights describe the conditions for scrutiny; they are not a safety verdict.
What METR can examine and publish
Anthropic says the embedded team can receive internal information, with exceptions for customer data, legally privileged material and model architecture. Highly relevant architecture details may be shared with explicit approval. Other METR staff are separated from this team’s work. Anthropic describes these access boundaries.
The company says it cannot veto METR’s publications, but can require removal of information in predefined categories. METR can publicly judge whether those redactions are justified without disclosing the withheld material. Publication freedom therefore coexists with confidentiality limits.
These are not universal terms for every evaluator. Anthropic describes three tiers: narrow, approval-dependent work; broader access with publication freedom; and METR-style embedding. The tiers have different access and publication conditions.
Nor does October 15 mark the beginning of all privileged access. METR’s September 22 assessment of Claude Opus 5.5 already referred to a separate team with elevated access. The news is the now-described arrangement; the reviewed sources do not establish when each evaluator started.
Publication independence has more than one stage
METR’s earlier work illustrates why the details matter. In its pilot report published in May, participating companies approved nonpublic material for inclusion but did not have approval rights over the final report. Those are separate decisions: what evidence can be disclosed, and whether the finished analysis can appear.
The practical implication is to examine both stages. Permission to publish a conclusion does not necessarily mean readers can inspect all its supporting evidence.
In its October 14 explanation of its approach, METR says it will disclose the scope and duration of relevant work, access limitations, consequential redactions and relevant conflicts of interest. This is its general policy, not a substitute for the terms and disclosures attached to a particular assessment.
What an assessment can establish
METR describes its role as gathering public evidence about the risk of humans losing control of AI agents, rather than deciding how companies should develop or deploy them. That remit helps explain why an evaluation partnership is not a product certification.
Its September Opus 5.5 assessment provides a concrete example. METR tested five tasks during ten business days of access through the model’s programming interface. The work concerned AI research capabilities; it did not assess compliance with specific policy thresholds or whether the model had particular alignment properties—how its behavior relates to intended goals. Those were METR’s tests, not new findings from this announcement.
Three questions to ask before relying on a report
For a purchasing or deployment decision, request the actual assessment and its process disclosures. The AI Evaluator Forum’s AEF-1 framework separates access, analytic independence and transparency. The following is a reader’s guide informed by that framework, not a score assigned to Anthropic or METR.
Check | Question for the report |
|---|---|
Access and resources | Which systems and information could evaluators inspect, for how long, and what remained outside their access? |
Analysis and publication | Who set the scope, who controlled the conclusions, and could redactions or approval requirements affect what readers see? |
Relevance to your use | What tasks and risks were actually assessed, what was excluded, and does that evidence address the decision you are making? |
Our recommendation is to keep the last question separate from the first two. Extensive access and strong publication rights can make an assessment more informative without making it answer every safety question. Evidence about AI research tasks, for example, does not establish how a customer-service assistant will handle your own records and permissions.
Application owners also need evidence from their own workflows. Our guide to Claude task tracing covers the separate question of investigating slow or failed tasks; monitoring an application is not the same as evaluating its model developer.
The full contract, termination provisions, dispute process and report schedule were not established by the reviewed sources. Nor did those sources independently demonstrate that the announced access works as described in practice. The next useful evidence will be the assessments themselves, with their methods, scope and limitations visible.
Methodology: AI-assisted reporting and analysis based on published company and evaluator statements, the AI Evaluator Forum framework and Reuters coverage, checked October 15, 2026. No model testing or private-contract inspection was performed. Practical recommendations are analysis, not tested outcomes.