Anthropic says six METR evaluators have deeper internal access

From a proposal to an onboarded team

Anthropic said on October 15 that it had onboarded six employees from METR, the nonprofit AI evaluator, with company equipment, office space and access to internal systems. It describes their access as roughly comparable to that of its own risk-assessment teams.

This is a material step beyond the plans covered in our September 19 report on Anthropic and Accenture. The new arrangement concerns a different team. Its terms should not be assumed to apply to every evaluator working with the company.

Closer access still has boundaries

Anthropic calls the arrangement employee-like access. It excludes categories such as customer data, legally privileged material and model-architecture information, although its footnote allows architecture disclosures when highly relevant and internally approved.

That distinction is important. Being inside a lab can reveal decisions and records that a public model test cannot. It does not mean every internal record is available, or that the evaluator has the authority of a regulator.

Who controls what reaches the public?

The company says METR can publish assessments without an Anthropic veto. Narrow, predefined redaction categories remain, and METR can publicly discuss whether it considers those redactions justified without revealing the protected information.

METR's October 14 proposal sets a wider context: agreed scope, private staff interviews and public reporting on access limits and redactions. It explicitly says this general proposal does not require fully equivalent employee-level access. That proposal is not the contract governing every specific engagement.

Access and disclosure therefore need to be read together. A review can be valuable even when information is withheld, but readers need to know whether missing evidence limits its conclusions.

Evidence, rather than a seal of approval

In a separate account of its approach, METR says it aims to explain the constraints, incentives and methods behind its findings. It commits to financial and operational independence and disclosure of relevant conflicts. Those are stated safeguards, not an independent audit of how this engagement is operating.

METR also says it does not intend to decide how companies develop or deploy their agents. Its proposals seek evidence about alignment, internal controls and AI-assisted research; they do not provide confident assurance that development is safe.

The next meaningful output is the assessment itself: what was examined, what the team could not see, what was removed and what the evidence supports. The new announcement establishes a reported access arrangement, not the result of that examination.