The development
What happened
Anthropic's October 15 announcement describes internal access for six METR evaluators assessing its frontier AI models and safety practices. The package includes hardware, office space and access comparable to its risk teams. METR can publish assessments without a company veto, but disclosure remains bounded by confidentiality categories.
Key facts
Customer data, privileged legal material and model architecture information are excluded categories; some architecture access can receive internal approval.
Anthropic can require redactions within predefined categories and ultimately determines category membership. METR may comment on access and redaction decisions.
Other evaluator partnerships have narrower information access and publication rights; discussions with Hugging Face are preliminary.
An October 16 clarification concerns METR's terms; the original announcement date remains October 15.
The context
Why it matters
Embedded access may allow closer examination of internal safety decisions. Independence must still be assessed against confidentiality controls. This voluntary arrangement is not regulatory certification or evidence that a model has passed a safety assessment.