Anthropic announced on October 15, 2026 that six researchers from METR, the nonprofit that evaluates frontier AI models, have moved into its offices with access roughly on par with its own risk assessment teams: company-issued laptops, internal tools, and colleagues encouraged to share information with them as they would with a fellow employee. This is not a one-off sample sent out for testing; the evaluators are sitting inside the lab for the long term.
Evaluators seated inside the lab
Anthropic compared the arrangement in its official blog post to banking supervision, where regulators keep resident examiners inside the largest banks. The stated goal is to close the information gap: outside evaluators have historically seen only near-final models, with limited visibility into training, internal incident records, or how risk judgments were made. Giving METR internal tools and a place in the daily flow of information moves evaluation from black-box sampling to on-site review.
Publication rights matter even more. METR can publish its own assessments of Anthropic's models, safety practices, and risk management, and Anthropic cannot veto publication. The company may require redactions only in narrow, pre-defined categories such as customer data, attorney-client privileged material, and model architecture details — and METR may publicly characterize whether it considers those redactions justified. Redactions themselves become part of the record rather than disappearing silently.
Three tiers, very different access
METR sits at the top of a three-tier external evaluation system Anthropic is building. The bottom tier runs narrow, scoped reviews and needs Anthropic's approval to publish; biosecurity nonprofit SecureBio, for example, is reviewing the chemical and biological weapons section of the company's August 2026 risk report. The middle tier receives more information and can publish without approval; Accenture's AI unit Faculty handles broader evaluations and adversarial testing there. Only the top tier, METR's, combines hardware, office space, internal systems, and guaranteed publication rights.
The move delivers on a proposal CEO Dario Amodei made in a September 2026 essay: embed outside evaluators inside labs as a first step toward slowing uncontrolled risk. Incidents of agents acting outside their instructions have surfaced every week or two since late August, sharpening calls for independent checks; the recent restoration of the UK AI Security Institute's pre-release testing access shows that evaluation access itself is becoming contested ground, as covered in how the UK AISI regained testing access.
What can be checked, and what stays closed
Access has limits. Customer data, legally privileged information, and model architecture details remain off-limits to METR, and Anthropic still chooses who its evaluators are. Reuters had previously reported that evaluators and labs spent weeks negotiating over publication control, with Anthropic becoming more open only recently. The company says it is in talks with additional organizations, but for now the three tiers are the whole roster.
The significance is not six desks. Anthropic has set a reference ceiling for what independent evaluation can mean: embedded presence, internal tools, and publication that cannot be blocked. Other frontier labs have committed to independent review under a White House accord; the open question is whether a second company opens the same door — and when METR's first embedded assessment appears, and what it is allowed to say.
