Michael JohnsonExternal AI risk assessment depends on what information researchers can access: companies usually know more about internal tests, model training, and safeguards than outside evaluators do. On October 15, 2026, Anthropic said it had begun giving the METR team access to its internal tools and systems. Six staff members of the nonprofit received company equipment and workspaces; METR will be able to publish its risk assessment findings independently.
According to Anthropic’s description of the agreement, the group’s access is broadly comparable to that of the company’s internal risk assessment teams. Anthropic employees have been asked to share information with METR participants as they would with colleagues working on risk assessment. This allows METR to examine not only model behavior but also information about training, evaluation processes, and safeguards.
What METR can examine
METR will be able to publish analyses of Anthropic’s risk levels, incidents, and practices. The company says it cannot veto publication, but can request redactions of information in predefined categories—for example, customer data, material protected by attorney-client privilege, and information about model architecture. Anthropic retains the decision on whether a particular passage falls under these exceptions. METR, in turn, retains the right to describe the level of access it received and the information provided.
That means the independent assessment will combine access to internal materials with limits on disclosing certain information. Readers will need to consider both METR’s findings and its description of which materials the group could examine and what restrictions affected publication. Anthropic says the collaboration is intended to narrow the gap between what outside evaluators know and what the lab itself knows.
How the access tier works
Anthropic calls its agreement with METR the third tier in its evaluator access framework. It provides access comparable to that of internal risk assessment teams and publication rights similar to those of the second tier. The work is limited to a small group at METR operating under binding agreements; the organization’s other staff, including those working on other Anthropic projects, are kept separate from this work.
The company said it has worked with METR for about a year, including on reviews of risk reports. Other partnerships are at earlier stages: SecureBio is reviewing the chapter on chemical and biological weapons in Anthropic’s August report, and the company is considering expanding access after that work is complete. Anthropic is developing separate collaboration with Accenture and Faculty on model evaluation, red teaming, and safeguards testing. The company is in preliminary discussions with other organizations, including Hugging Face.
What this means for companies
For organizations deploying models in sensitive processes, the practical value of the framework will be shown by the published assessments: which risks were tested, what data was examined, and what limitations affected the findings. Anthropic’s announcement describes the terms of access and publication but does not report new model assessment results. Corporate buyers should therefore treat future METR materials as evidence about specific evaluations, not as a blanket conclusion about the safety of all the company’s models.
The October 15 announcement changes outside evaluators’ access to internal information and gives METR the ability to publish its findings subject to specified exceptions. The content of future assessments, and the clarity with which their conditions are described, will show how useful this framework is for independent AI risk evaluation.