Franklin AI News Brief

Anthropic gives embedded METR evaluators deeper access and publication rights

Key Takeaways

  • Anthropic has onboarded six METR staff with employee-like access, while preserving defined information exclusions and allowing evaluators to discuss required redactions.
  • Anthropic has begun embedding a team from independent AI evaluator METR inside its operations.
  • In its announcement about external evaluation, the company describes the access provided, the evaluator's publication rights and the information that remains outside the arrangement.
  • Give evaluators access beyond the model interface Six METR employees have been onboarded.
  • Anthropic says the team has company hardware, office space and internal tool access roughly comparable to its own risk-assessment teams.

Anthropic has begun embedding a team from independent AI evaluator METR inside its operations. In its announcement about external evaluation, the company describes the access provided, the evaluator's publication rights and the information that remains outside the arrangement.

Give evaluators access beyond the model interface

Six METR employees have been onboarded. Anthropic says the team has company hardware, office space and internal tool access roughly comparable to its own risk-assessment teams.

The arrangement has defined exclusions, including customer data, attorney-client privileged material and information about model architectures. “Employee-like access” therefore describes a specific risk-evaluation arrangement, rather than unrestricted access to everything the company holds.

Other METR staff are separated from the embedded team's work under enforceable agreements.

Distinguish publication rights from unrestricted disclosure

Anthropic cannot veto METR's publication of its analysis. It can require redaction within narrow, predefined information categories, while METR can publicly characterize those redactions and explain whether it considers them justified.

That provides a mechanism for readers to hear about a disagreement over disclosure. It does not remove the confidentiality restrictions themselves.

The company says METR will publish assessments of risks, incidents or practices. The announcement establishes the arrangement, rather than reporting a new independent finding that Anthropic's models have passed a particular assessment.

Compare the different evaluator tiers

Anthropic describes three access tiers. The first supports narrowly scoped investigations with publication requiring company approval. The second permits publication without that approval, subject to specified redactions. The third adds employee-like access, as in the METR arrangement.

Other partnerships remain at earlier stages. Anthropic names SecureBio and a developing partnership involving Accenture and Faculty, with more limited information and publication rights in early engagements.

The practical test of the arrangement will be the published assessments and explanations of any withheld material. Access and reporting terms matter because they shape what an outside evaluation can establish.

Our read

Franklin AI Take

Anthropic cannot veto METR's publication of its analysis. Readers should still examine the eventual report's scope and any disclosed redactions, since publication rights and complete information access are separate conditions. The strongest evidence of this arrangement's value will come from what evaluators can explain about actual findings and disagreements.