Anthropic Gives METR Employee-Like Access to Evaluate Its AI Models
Anthropic says it has onboarded six METR employees with company hardware and internal access to independently assess risks from its AI models.
한 줄 답
What did Anthropic announce about independent evaluators of its AI models?
Anthropic says it has given a team of six employees from nonprofit evaluator METR employee-like access, including company hardware and internal tool access, to assess risks in its models. METR will publish its findings, and Anthropic says it cannot veto publication, though it can require redactions in narrow, predefined categories.
핵심 요약
- Anthropic says it has onboarded six METR employees with company hardware, office space, and internal tool access comparable to its own risk assessment teams.
- Anthropic says METR will publish its analysis of Anthropic's risk levels, incidents, or practices, and that Anthropic cannot veto publication but can require redaction of information in narrow categories such as customer data or model architecture details.
- Anthropic says it is also working with SecureBio and Accenture as evaluators and is in early discussions with other organizations, including Hugging Face.
- Anthropic describes three tiers of evaluator access, ranging from narrowly scoped investigations requiring company approval to publish, up to the employee-like arrangement it has with METR.
- Anthropic says SecureBio is currently reviewing the chemical and biological weapons chapter of its August Risk Report, after previously completing and publishing a pilot review.
Anthropic says it has begun giving a team from METR, a nonprofit that evaluates frontier AI models, "employee-like access" to its systems as part of an expanded push toward independent, third-party assessment of its models and safety practices.
According to Anthropic, it onboarded six METR employees with company hardware, office space, and access to internal tools and systems that the company says is roughly on par with what its own risk assessment teams have. Anthropic says its employees are encouraged to share information with this METR team as they would with a fellow employee working on alignment risk, with exceptions for a small number of defined categories such as customer data, attorney-client privileged information, and details about model architecture.
Anthropic says METR will publish its analysis of Anthropic's risk levels, incidents, or practices, and that Anthropic cannot veto that publication. The company says it can require redactions only in the narrow categories it has defined, and that METR remains free to publicly characterize those redactions and state whether it believes they were justified, without revealing the redacted information itself.
Anthropic says its engagement with METR is limited to a small team working under enforceable agreements, and that other METR employees, including those who work with Anthropic on separate projects, are siloed from that team's work.
Other evaluator partnerships
Anthropic says it is also working with SecureBio and Accenture as evaluators, and is in early discussions with several other organizations, including Hugging Face. The company says these other partnerships are at earlier stages, with more limited information access and more limited publication rights than the METR arrangement.
Anthropic describes three tiers of evaluator access it is developing. In the first tier, the company says, evaluators conduct narrowly scoped risk investigations, Anthropic decides what access to provide, and evaluators can publish findings only with Anthropic's approval. In the second tier, evaluators get more information and can publish without approval, though Anthropic can still require removal of specific categories of information. The third tier, which Anthropic says matches its current arrangement with METR, combines employee-like access with publication rights similar to the second tier.
Anthropic says SecureBio is currently reviewing the chemical and biological weapons chapter of its August Risk Report, after completing and publishing a pilot review of a previous risk report, and that Anthropic will consider moving to projects with increased access once that review is complete. With Accenture and Faculty, Anthropic says it is working toward a partnership that includes evaluating and red-teaming models, conducting alignment assessments, testing model safeguards, and building capabilities across the organizations involved.
Anthropic says it worked with METR on several reviews of its Risk Reports over the past year, and that last month it committed to stronger independent third-party assessment of its models and signed the White House Accord on Super Intelligence, in which signatory companies pledged to partner with independent external auditors or evaluators.
Anthropic says it is embedding evaluators voluntarily and believes every frontier AI company should do the same.
Source: Anthropic, "An update on independent evaluators," published October 15, 2026.
자주 묻는 질문
- What access did Anthropic give METR?
- Anthropic says it onboarded six METR employees with company hardware, office space, and access to internal tools roughly on par with its own risk assessment teams, with exceptions for categories like customer data and model architecture information.
- Can Anthropic block METR from publishing its findings?
- Anthropic says it cannot veto publication but can require METR to redact information that falls into narrow, predefined categories, and METR can publicly state whether it believes those redactions were justified.
- Which other organizations is Anthropic working with on evaluation?
- Anthropic says it is working with SecureBio and Accenture, alongside Faculty on the Accenture partnership, and is in early discussions with other organizations including Hugging Face.
- What are Anthropic's three tiers of evaluator access?
- Anthropic describes a first tier with narrowly scoped investigations requiring company approval to publish, a second tier with broader access and no publication approval needed, and a third tier with employee-like access, which is the arrangement it has with METR.