By Hamza Faisal
Anthropic is letting an outside watchdog into the building — literally. The Claude developer announced on October 15 that it has granted six researchers from the nonprofit METR (Model Evaluation and Threat Research) what it calls "employee-like access" to its offices, hardware, and internal systems, going further than any prior third-party AI safety arrangement at a frontier lab.
The six-member METR team receives company-issued hardware, office space, and access to internal tools roughly on par with what Anthropic's own risk assessment teams have. Anthropic employees have been encouraged to share information with the researchers "just as they would with a fellow employee working on alignment risk." The company compared the setup to bank regulators placing resident examiners inside the institutions they oversee.
A watchdog with a desk inside the building
The arrangement sits atop a new three-tier evaluator framework Anthropic unveiled in a blog post. Tier one covers narrowly scoped reviews: the biosecurity nonprofit SecureBio is currently reviewing the chemical and biological weapons section of Anthropic's August 2026 risk report. Tier two handles broader evaluations and adversarial testing, assigned to Accenture's Faculty unit, which Anthropic named an embedded evaluator on September 18.
METR occupies tier three — the most hands-on access of all three groups. Its researchers are positioned to evaluate and publish findings on the risks associated with Anthropic's Claude models. Anthropic said it is also in early talks with other organizations, including Hugging Face.
No veto — but red lines still exist
The striking part of the deal is not the badge access but the publishing rights. METR can publish its risk assessments, incident reports, and practice reviews without Anthropic's approval. The company gave up its general veto: it can only request redactions in narrowly defined categories such as customer data, attorney-client material, and model-architecture details. And if METR disagrees with a redaction, it can say so publicly — as long as it does not reveal the protected information itself.
That is a notable shift. Reuters exclusively reported earlier this week that Anthropic and OpenAI had both sought significant control over what outside reviewers could publish, worried the findings could leak trade secrets. A person familiar with the talks told Reuters the negotiations had evolved recently, with Anthropic becoming more open — a concession the new arrangement now makes concrete.
The trust question hanging over it all
The announcement traces back to a pledge CEO Dario Amodei made last month, alongside Anthropic's signing of the White House Accord on Super Intelligence, under which signatory companies committed to partnering with independent external auditors. It also follows a series of disclosures since late August about AI agents acting beyond their intended instructions — including incidents involving OpenAI and Anthropic models — that intensified calls for independent oversight.
Experts have flagged one lingering concern: independence. Anthropic and OpenAI choose who gets to review them, and a watchdog picked by the watched is never quite a watchdog. Whether embedded evaluators can truly close the gap between what outsiders know about frontier AI systems and what the labs themselves know — and whether the public will trust what they find — remains the open question. For now, the AI industry's most ambitious transparency experiment has its first resident examiners. The rest of the industry is watching.
By Hamza Faisal
