Anthropic said in a company blog post on Nov. 4 that it has granted six researchers from METR, a nonprofit that tests frontier AI models, employee-level access to independently evaluate the company’s models, safety practices, and risk management.
Scope of access and safeguards
The METR team will use Anthropic-issued laptops, office space, and internal tools described as comparable to those used by Anthropic’s own risk assessment teams. Anthropic said employees are encouraged to share information with METR “just as they would with a fellow employee working on alignment risk.”
Anthropic compared the arrangement to the banking practice of placing resident examiners inside large banks. The company framed the move as an effort to “be publicly accountable for the decisions we’re making” and to narrow the information gap between external evaluators and company insiders.
Publication rights and redaction rules
METR will publish its own assessments of Anthropic’s risks, incidents, and practices, and Anthropic has relinquished the right to veto those publications. Anthropic said it may require redactions only in narrow, predefined categories such as customer data, attorney-client privileged information, or details about model architectures. METR is allowed to publicly describe redactions and state whether it believes they were justified, provided the redacted information itself is not revealed.
Anthropic also described a three-tier evaluation system. METR occupies the top tier with employee-level access. Evaluators in the middle tier receive more information and can publish without approval, while lower-tier reviewers like biosecurity nonprofit SecureBio conduct narrower reviews and require Anthropic’s approval to publish.
The announcement fulfilled a proposal made by Anthropic CEO Dario Amodei in a September essay advocating for embedding outside evaluators inside AI labs. OpenAI quickly made a similar pledge, and Anthropic, OpenAI, Google, Meta, Nvidia, and xAI have signed a White House accord committing to independent review.
Anthropic framed the arrangement as a response to a series of recent incidents involving AI agents acting outside instructions, which the company said have prompted calls for more independent checks since late August.
Original source: TechCrunch AI