Anthropic Grants Outside AI Watchdog METR "Employee-Like Access"

Anthropic has granted external evaluators "employee-like access" to assess the risks of its models, following through on a commitment to increase safeguards over concerns the technology could run catastrophically out of control.

The evaluator is METR, a nonprofit that has already worked with leading labs including OpenAI, Anthropic, Google and Meta. METR now has a team embedded in the Claude maker's office with access to its hardware and internal systems.

The arrangement is unusual in an industry that has historically kept safety testing close to home. Labs typically publish model cards, commission outside red-teaming or submit to voluntary evaluations after release. Embedding an independent organization inside the company, with the same practical access an employee would have, moves third-party scrutiny upstream into the development process itself.

Why does embedded access matter?

A greater role for third-party evaluators is among the main concrete steps the labs have proposed in response to rising concern about the dangers more capable AI systems could pose to the economy, cyber security and humanity at large.

The distinction matters because of what external researchers have historically lacked. Outside audits conducted through application programming interfaces or staged evaluations can miss behaviors that only appear under full internal conditions: access to raw compute, unreleased checkpoints, internal tooling and the scaffolding around a model. "Employee-like access" closes that gap by design rather than by negotiated exception.

For Anthropic, the move follows through on a stated commitment. The company positioned the arrangement as a safeguard against the scenario its own leadership has flagged repeatedly: that sufficiently capable systems could run catastrophically out of control.

Who is METR?

METR is a nonprofit evaluator, not a regulator. Its client list spans the largest AI developers: OpenAI, Anthropic, Google and Meta.

That track record is the basis for the trust the arrangement requires. An embedded team with access to hardware and internal systems could not function inside a lab it had never worked with. It also concentrates responsibility: if METR becomes one of the few organizations with inside access across multiple frontier labs, its methods and independence will face correspondingly sharp scrutiny.

The model resembles nothing in traditional software auditing. A financial auditor gets documents. METR gets a desk.

What does the arrangement change in practice?

Three things distinguish embedded evaluation from standard practice, based on what Anthropic and METR have described:

• Timing. Assessment happens during development, not only before or after release.
• Depth. The team works on Anthropic's hardware and internal systems, not through external interfaces.
• Standing. The access is "employee-like," meaning ongoing presence rather than a one-off audit.

Each element addresses a documented weakness of voluntary external testing. Post-hoc evaluations arrive after deployment; API-based testing sees only what the interface exposes; commissioned audits end when the contract ends.

What are the stakes for the wider AI debate?

The arrangement lands amid a broader argument over who should verify lab safety claims. Governments have moved slowly on mandatory audit regimes. In their absence, the major labs themselves have proposed third-party evaluation as one of the main concrete responses to public concern about economic disruption, cyber security threats and catastrophic risk.

Anthropic's move gives that proposal its most complete form to date at a major lab. If embedded access produces findings the public can weigh — and if METR's independence holds under the proximity — the arrangement could become the template other labs face pressure to copy. OpenAI, Google and Meta have all worked with METR already; the question is whether any of them follow Anthropic in letting the nonprofit through the front door.

The precedent cuts both ways. Access without publication changes little. The test ahead is whether "employee-like access" produces employee-like candor about what the evaluators find.