Dylan Di Filippo
Dylan Di FilippoAI journal

Six METR evaluators gain internal access at Anthropic

Anthropic / METRGovernance and safety

Six METR evaluators gain internal access at Anthropic

The arrangement goes beyond API testing while retaining exclusions and redaction rights.

External evaluation underway

Official visual published by Anthropic / METR for this announcement
Anthropic / METR — visuel officiel · Image source ↗

Verified facts

  • Six METR staff receive hardware, office space and tools similar to internal risk-assessment teams.
  • METR can publish without an Anthropic veto; predefined categories remain excluded, including customer data, privileged information and architectures.
  • Anthropic decides how information is classified for redaction, while METR can comment on its access level. SecureBio and Accenture have earlier-stage arrangements; Hugging Face is still in discussions.

What changes

The question becomes what auditors can actually see and publish.

Practical implication

Read METR’s future conclusions separately from the partnership announcement.

Availability

Access is restricted to the evaluators involved; this is neither a new user feature nor public access to internal systems.

Limits and uncertainty

The arrangement is described by Anthropic. Deep access and formal independence do not certify an absence of risk.

← Back to this weekly edition

A journal by Dylan Di Filippo. Facts, vendor claims and editorial analysis are kept separate. Coverage is not exhaustive.

Sources and methodology