Artificial Intelligence
Anthropic expands access for independent evaluators
Anthropic said it had begun giving METR employee-like access to its systems, while setting out a tiered model for other independent evaluators and limiting what they could see or publish.
- Anthropic said it had started giving a METR team employee-like access to its hardware, office space and internal tools.
- The company said METR would publish its risk assessments and Anthropic could only require redactions in narrow categories.
- Anthropic said it had worked with SecureBio on prior reviews and was in early discussions with other organisations, including Hugging Face.
- Anthropic said SecureBio was in the first tier and was reviewing the chemical and biological weapons chapter of its August Risk Report.
- Anthropic said it had signed the White House Accord on Super Intelligence and had committed to stronger third-party assessment of its models and safety practices.
Anthropic said last month that it had committed to stronger independent assessment of its models and safety practices, and it said it had signed the White House Accord on Super Intelligence. It then said it had begun giving a METR team employee-like access.
The company said METR would publish its risk assessments. Anthropic said it could demand redactions only for limited categories such as customer data, privileged legal material or model architecture details, and METR could publicly challenge those redactions.
Anthropic said six METR employees had been given company hardware, office space and access to internal tools comparable with those used by its own risk assessment teams. It said the arrangement was confined to a small team under enforceable agreements, with other METR staff kept separate.
Anthropic said it was also working with SecureBio, Accenture and Faculty, and had begun discussions with other organisations, including Hugging Face. It said SecureBio was in the first tier of its system and was reviewing the chemical and biological weapons chapter of its August Risk Report.
Named in this story
Companies
- Anthropic
- announced the evaluator access arrangement and its wider tiered system
- Accenture
- was working with Anthropic on an evaluator partnership
- Faculty
- was working with Anthropic on an evaluator partnership
- Hugging Face
- was mentioned as one of the organisations in early discussions with Anthropic
Organisations
- METR
- received employee-like access and will publish risk assessments
- White House Accord on Super Intelligence
- was signed by Anthropic as a pledge to work with independent auditors or evaluators
- SecureBio
- was working as a first-tier evaluator on a chapter of Anthropic’s August Risk Report
How the source tells it
The register was sober and corporate, with repeated attempts to reassure readers that the access programme remained bounded and responsible.
- Controlled reassurance uses a bank-supervision analogy and detailed access limits to frame the arrangement as safe
- Promotional advocacy presents its tiered evaluator model as a template other frontier AI companies should adopt