Anthropic gives METR an inside view of its risk work
Anthropic says six METR staff now have broad internal access and rights to publish risk assessments, subject to defined information exclusions and redactions.
Anthropic says it has equipped six METR evaluators with hardware, office space and internal access broadly comparable to its own risk teams.
The October 15 announcement excludes categories including customer data, privileged legal material and model architecture information. [1]
The significant provision is publication: Anthropic says it cannot veto METR’s assessments, although it can require narrowly defined redactions. METR may publicly contest those redactions without revealing the protected information. [1]
Access and publication rights could make outside scrutiny more informative. The test will be the assessments that emerge and how disagreements are handled. This is a voluntary arrangement described by Anthropic, not a new regulatory authority or a finding that its models are safe. The underlying agreement was not available in the announcement. [1]
Sources & details 1 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- Anthropic · Evaluator access and publication terms
Original October 15 company account, updated that day. No underlying agreement or completed assessment of the new arrangement was inspected.
Images & figures
AI-generated editorial illustration of independent access with boundaries and a route to publication; not documentary evidence or a measured diagram.
Euangelion · AI-generated editorial illustration · Source · Usage terms