Anthropic gives METR internal access to assess AI risks
Six external evaluators receive access similar to internal risk teams and can publish without a veto. The arrangement retains information exclusions and rules for redacting sensitive material.
Summary
Independent AI safety assessment becomes more informative when researchers can observe more than a model’s public responses. Anthropic announced on October 15 that six METR employees had received internal access similar to its own risk-assessment teams.
METR is a nonprofit that studies AI capabilities and risks. Its evaluators receive equipment, office space and access to internal tools and systems. They can publish analyses of the company’s risks, incidents and practices without Anthropic being able to veto publication.
In practice
This access allows evaluators to examine the context in which models are developed, assessed and used. That can include evidence unavailable in an external test: internal incidents, the operation of safeguards and information needed to interpret results.
There are defined limits. Customer data, attorney-client privileged information and model-architecture information are among the excluded categories. Anthropic can require redaction of information within those categories and ultimately determines what falls within them. METR reserves the right to make certain statements about the access it received.
Anthropic distinguishes three partnership tiers. At the most limited level, publication requires its approval; METR’s tier offers deeper access and publication freedom. External evaluators should therefore not be assumed to work under identical conditions.
Context
In its own account of its approach to risk assessment, published on October 14, METR argues that the public should understand access limitations, significant redactions and pressures that may affect conclusions. Its stated role is to provide evidence, rather than decide how companies should deploy agents.
The arrangement is voluntary. It is neither a safety certification nor a substitute for public oversight. Its usefulness will depend on what evaluators can investigate, publish and explain about the restrictions they encounter.
Why it matters
- Internal access can reveal risks that public testing cannot observe.
- Independence requires publication freedom and transparency about limits and conflicts of interest.
- The quality of published assessments will matter more than the partnership announcement alone.