Anthropic Gives METR Researchers Employee-Level Access to Internal Evaluations
Anthropic said six METR researchers will get hardware, office space and internal tool access, and can publish assessment reports the company cannot veto.
Anthropic announced on October 15, 2026 that it has granted six METR researchers access to its internal evaluation systems at a level roughly the same as its own risk assessment team, according to aitimes.com. The researchers will receive hardware, office space and access to Anthropic's internal tools and systems.
METR can independently publish assessment reports on Anthropic's risk levels, incidents and safety practices. The company cannot refuse publication and may only request deletion or revision in pre-defined categories, including customer data, attorney-client privileged information and model architecture.
Anthropic is preparing a three-tier system for external evaluators' access, with METR participating at tier 3. The company is also pursuing cooperation with Accenture and its AI business unit Faculty on model evaluation, red-team testing, alignment evaluation and safety guardrail verification, and said it is in talks with other organizations including Hugging Face.
Anthropic compared the arrangement to a bank supervision system.
Background
METR published proposals on October 14 to better inform the public about 'loss of control' risks in frontier AI development. In the previous month, Anthropic CEO Dario Amodei proposed involving independent evaluators inside companies to manage frontier AI development speed, after which major AI companies said through the White House's superintelligence agreement that they would cooperate with independent external evaluators. SecureBio is conducting chemical and biological weapon related risk assessments in Anthropic's August risk report.
Quick answers
What access did Anthropic give METR researchers?
Six METR researchers received hardware, office space and access to Anthropic's internal tools and systems, roughly the same level as Anthropic's own risk assessment team.
Can Anthropic block METR's reports?
No. Anthropic cannot refuse publication and can only request deletion or revision for pre-defined categories such as customer data, attorney-client privileged information and model architecture.
Who else is Anthropic working with on evaluation?
Anthropic is pursuing cooperation with Accenture and its AI business unit Faculty, and said it is in talks with other organizations including Hugging Face.