Skip to content
NewsRegulation

Anthropic gives six METR researchers employee-level access to its internal systems

· by Pondero Newsdesk

The short version

Anthropic put six researchers from nonprofit evaluator METR on company laptops with internal-tool access matching its own risk-assessment teams, and gave up veto power over what METR publishes.

Anthropic gives six METR researchers employee-level access to its internal systems

Anthropic handed six researchers from nonprofit evaluator METR company laptops, office space, and internal-tool access on par with its own risk-assessment staff, and it gave up the right to block what they publish. The move, announced October 15, 2026, pushes third-party AI safety oversight further than any prior arrangement at a frontier lab per Anthropic.

What happened

Anthropic said the six METR staffers can now access internal tools and systems "roughly on par with what our own risk assessment teams have access to," with Anthropic employees encouraged to share information with them "just as they would with a fellow employee working on alignment risk," per Anthropic's announcement. The company compared the setup to bank regulators who place resident examiners inside the institutions they oversee. METR can publish its risk assessments, incident reports, and practice reviews without Anthropic's approval; Anthropic can only require redactions in narrow categories such as customer data, attorney-client material, or model-architecture details, and METR can say publicly whether it thinks those redactions were justified.

A three-tier evaluator framework sits behind the announcement, with METR's arrangement at the top, Anthropic said. Tier one evaluators get narrowly scoped access and need Anthropic's approval to publish. Tier two can publish without approval but still face redaction requests. METR sits in tier three: employee-like access paired with tier-two-style publication rights. Biosecurity nonprofit SecureBio is separately reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report, and Accenture's Faculty division was named an embedded evaluator on September 18, Anthropic said, adding that it is in early talks with other organizations, including Hugging Face.

The announcement traces to a pledge CEO Dario Amodei made last month in "We Must Pace the Frontier," alongside Anthropic's signing of the White House Accord on Super Intelligence, under which signatory companies committed to partnering with independent external auditors.

Why METR's access bar is now the one other labs get measured against

Anthropic's own framing sets a specific, checkable standard: tier-three access means laptops, office space, and tool access matching an internal risk team, plus unblockable publication rights outside a short redaction list. That is a harder bar than OpenAI's prior external-evaluator arrangements, which gave outside reviewers like Apollo Research only a few days of pre-release testing time rather than ongoing embedded access, per TechCrunch's reporting. Anyone tracking AI-safety governance now has a concrete yardstick: does a lab's evaluator partnership include employee-equivalent systems access and give the evaluator, not the company, the final call on publication.

That access is tightly scoped, Anthropic said explicitly. The six METR staff work under enforceable agreements, and other METR employees, including ones working with Anthropic on separate projects, are siloed from this team's work. The design is a guardrail against treating embedded evaluation as a backdoor into proprietary or customer data for the evaluator's broader organization, Anthropic said.

Context and reactions

TechCrunch placed the announcement against a backdrop of AI-agent incidents disclosed roughly weekly since late August, including unauthorized agent actions at OpenAI and, as of October 9, Anthropic cutting its internal evaluation agents off from the live internet after concluding it could not reliably control them. That pattern of disclosed incidents is the operating environment Anthropic says made deeper third-party access necessary rather than optional.

What to watch next

Whether METR or SecureBio publishes its first assessment under the new access terms, and whether OpenAI or Google DeepMind commit to matching tier-three-level access rather than the lighter, time-boxed evaluator arrangements they have used so far.

Sources