Anthropic Gives METR Researchers Employee-Level Access

Anthropic has opened its doors to an outside safety team. Six researchers from METR, a nonprofit that tests frontier AI models, now have employee-level access to the company. Their job is to independently assess Anthropic's models, its safety practices and how it manages risk. The company announced the arrangement in a blog post on Thursday.

The timing matters. Since late August, a new incident of an AI agent acting outside its instructions has been disclosed roughly every week or two. One recent case involved an Anthropic model sending a false murder tip to police. These episodes have strengthened calls for independent checks on AI companies, and Anthropic is now trying to answer them with more than promises.

What "employee-level" means in practice

The METR team works much like internal staff. Its researchers have company laptops issued by Anthropic, office space, and access to internal tools. According to Anthropic, that access is roughly on par with what its own risk assessment teams use.

Access goes beyond hardware and software. Anthropic says its staff are encouraged to share information with the METR researchers "just as they would with a fellow employee working on alignment risk." Alignment, in this context, refers to making sure AI systems behave in line with their developers' intentions.

The company stated its goal directly: "We want to be publicly accountable for the decisions we're making." It also said it wants to narrow the gap between what outside evaluators know about a frontier AI company's systems and what the company knows itself.

To explain the setup, Anthropic pointed to banking. Regulators place resident examiners inside the largest banks, and Anthropic says the METR arrangement follows a similar logic.

Publication without a veto

The most notable part of the deal concerns what METR can publish. The nonprofit will release its own assessments of Anthropic's risks, incidents and practices, and Anthropic has given up the right to block them.

There are limits. "Anthropic can't veto publication, but we can require them to redact information if it falls into narrow, pre-defined categories," the company said. Those categories include customer data, material covered by attorney-client privilege, and information about its model architectures.

METR can also speak publicly about those redactions. Anthropic said the nonprofit is free to describe them and say whether it thinks they were justified, as long as it doesn't reveal the redacted information itself.

A three-tier system

METR sits at the top of a three-tier structure that Anthropic is building for outside evaluators. So far, it is the only organization at that level.

• Bottom tier: Evaluators run narrow reviews and need Anthropic's approval before publishing. Biosecurity nonprofit SecureBio is one example.
• Middle tier: Evaluators receive more information and can publish without approval.
• Top tier: Evaluators get everything the middle tier gets, plus METR's employee-level access.

Delivering on a September pledge

Thursday's move follows through on a promise from Anthropic CEO Dario Amodei. In a September essay, he proposed embedding outside evaluators inside AI labs as a first step toward slowing the technology's development.

OpenAI quickly made the same pledge. Since then, both companies have signed a White House accord committing to independent review, along with Google, Meta, Nvidia and xAI.

Our Take

The core idea here is simple: outside reviewers learn more when they work inside. Most third-party AI evaluations so far have depended on whatever access and documentation a company chose to provide. Giving METR the same tools as internal risk teams, and removing the publication veto, changes that relationship in a meaningful way, at least on paper.

The arrangement also fits a pattern. Anthropic has been tightening its own safeguards for weeks, including a decision to cut its internal AI evals off from the live internet. Across the industry, labs are acknowledging that their systems are not fully under control. OpenAI, for instance, recently signaled that its flagship model needs more alignment work. Independent oversight looks like the next logical step.

Still, much depends on details that aren't visible yet. Six researchers is a small team for a company of Anthropic's size. The redaction categories sound narrow, but how often they get used will matter. METR's freedom to comment on redactions could become the most useful safeguard in the whole setup.

It is worth watching three things. First, what METR's first published assessment actually says, and how heavily it is redacted. Second, whether more evaluators move into the top tier. Third, whether OpenAI, Google, Meta, Nvidia and xAI offer comparable access under the White House accord, or settle for lighter forms of review. The banking comparison suggests a lasting oversight model, but that will only hold if the access proves real and the reports are allowed to be critical.