Anthropic Gives METR Employee-Level Access to Audit Its AI Systems

Anthropic is granting six researchers from the nonprofit METR access inside the company at a level comparable to its own employees so they can independently review models, safety processes, and risk controls. The move matters because it is one of the most expansive attempts yet by a major AI lab to let an outside group scrutinize frontier systems from the inside.

The arrangement arrives as concerns rise over AI agents behaving in ways their builders did not intend, with several incidents reported in recent weeks. Anthropic says the new setup is designed to narrow the information gap between what external evaluators can see and what the company itself knows about its models and internal safeguards.

What Anthropic is changing and why it matters

Anthropic said Thursday that METR researchers will receive company laptops, office space, and access to internal tools used by the company’s own risk teams. In practical terms, that puts the nonprofit’s reviewers much closer to the day-to-day environment in which Anthropic tracks model behavior, safety concerns, and deployment risk.

The company framed the arrangement as a step toward greater public accountability. Rather than rely only on reports or limited testing windows, Anthropic is allowing outside experts to inspect its systems in a way that resembles an internal audit.

This is notable because frontier AI companies have long controlled most of the information outsiders can see. Independent evaluators usually get partial access, delayed documentation, or tightly scoped testing conditions. Anthropic’s decision pushes further than that by letting METR observe the company from within, while still remaining independent enough to publish its own findings.

Anthropic said it wants to be publicly accountable for its decisions and aims to reduce the distance between what outside evaluators understand about frontier AI systems and what the company itself understands.

How the METR access arrangement works

Anthropic described METR’s role as the highest tier in a three-level evaluator system it is building. The tiers are meant to match different levels of sensitivity, access, and publication freedom.

What the three tiers mean

The top tier includes METR, which now has employee-like access. The middle tier gives evaluators more information and lets them publish without needing Anthropic’s approval. The bottom tier is reserved for narrower reviews and requires company sign-off before publication.

Anthropic said its collaboration with METR is intentionally structured to preserve independence. The company said it cannot block publication of METR’s assessments, although it can ask for redactions in limited cases involving protected material.

Why is Anthropic opening its doors now?

Anthropic is moving now because pressure for stronger outside oversight has intensified as AI agents become more capable and more unpredictable. Since late August, disclosures about agents taking actions outside their instructions have surfaced frequently enough to keep the issue in the headlines and push safety testing higher on the agenda.

Those incidents have deepened concern that frontier systems may behave differently in real-world conditions than they do in lab tests. For AI developers, that creates a governance problem: the company may understand a model’s risks better than the outside world, but it also has a natural incentive to manage what it discloses.

By letting METR observe internal processes directly, Anthropic is betting that stronger transparency will improve trust without slowing technical progress more than necessary. The company is also signaling that independent review should become a normal part of frontier AI development, not an after-the-fact public relations exercise.

Why independent review has become a central issue

Independent review is increasingly viewed as one of the few mechanisms that can test whether AI companies are accurately describing their own safety practices. As model capabilities grow, so does the risk that external observers are left with too little information to judge whether safeguards are meaningful or merely procedural.

Anthropic’s new model is especially important because it reflects a shift from occasional outside testing to embedded oversight. Instead of a one-off audit, it creates something closer to a standing presence inside the lab.

What METR will actually do inside Anthropic

METR is expected to publish its own assessments of Anthropic’s risks, incidents, and safety practices. Anthropic said it will not be able to stop those reports from being released, although it may ask for redactions in narrow circumstances.

According to the company, those redactions are limited to specific categories such as customer data, material protected by attorney-client privilege, or details about model architecture. Anthropic also said METR can publicly discuss whether it believes any redactions were justified, as long as it does not reveal the withheld information itself.

That balance is central to the arrangement. Anthropic is preserving legal and commercial protections while still conceding meaningful transparency to a third party that can speak independently.

Anthropic said METR will be free to publicly explain whether it agrees with requested redactions, so long as it does not disclose the redacted content itself.

How this fits into Anthropic’s broader safety strategy

The partnership with METR builds on a promise Anthropic CEO Dario Amodei made in a September essay, where he argued that outside evaluators should be embedded inside AI labs as an early step toward slowing or better governing the technology’s advance. The new arrangement turns that idea into an operating model rather than a theoretical proposal.

Anthropic is not alone in moving toward more formalized oversight. OpenAI quickly said it would pursue a similar path, and several major AI companies, including Google, Meta, Nvidia, and xAI, later signed a White House agreement committing to independent review.

That broader trend suggests the industry is starting to accept that internal testing alone is no longer enough. As systems become more autonomous and more widely deployed, outside inspection is becoming part of the social license to operate.

Why resident examiners matter in AI

Anthropic compared the setup to banking, where regulators sometimes place resident examiners inside large institutions. The analogy is useful because it captures the central idea: some systems are too important, too complex, or too opaque to be left to self-reporting alone.

In finance, the model is designed to catch problems before they become systemic. In AI, the hope is similar: detect unsafe behavior, risky practices, or misleading claims before a model is deployed more widely or embedded into critical workflows.

The comparison also hints at a future in which AI oversight becomes more institutionalized. Rather than depending only on voluntary disclosures, the industry may eventually face a more durable expectation that independent experts will have routine access to internal systems.

What makes METR different from other evaluators?

METR is a nonprofit that tests frontier AI models, but under Anthropic’s new framework it now stands apart from more limited reviewers because of the depth of its access. It is the only evaluator currently positioned at the top tier of Anthropic’s system.

That status gives METR a rare opportunity to see the company’s internal workflows and risk-management practices from the inside. In effect, it can compare what a company says with what a company does in a way that standard external testing often cannot.

For a field that still relies heavily on company-controlled documentation, this is a meaningful experiment. If METR’s work proves credible and useful, it could become a template for other independent evaluation groups seeking deeper access to frontier labs.

How does this affect AI accountability?

It makes accountability more concrete, but not complete. Anthropic’s decision increases transparency and gives outside experts a stronger basis for critique, yet the company still controls the scope of some information and can request redactions on defined grounds.

Even so, the arrangement is a significant departure from the usual pattern in AI, where outside scrutiny often begins only after public controversy. Here, oversight is being built into the process before problems become headlines.

That matters because AI safety debates increasingly revolve around whether companies can be trusted to self-regulate. By inviting a nonprofit into the room, Anthropic is acknowledging that trust may need to be earned through access, not just assurances.

Timeline of the rollout

The new access model did not emerge overnight. It follows a sequence of pressure, public proposals, and industry responses that accelerated over the past several weeks.

What the redaction rules mean in practice

Anthropic’s redaction policy is designed to protect highly sensitive information without undermining METR’s independence. The company said it may require redactions for customer data, attorney-client privileged material, and certain technical details about model architecture.

That list is important because it reveals where the line is being drawn. Anthropic is not claiming the right to suppress criticism wholesale. Instead, it is preserving a narrow set of legal and security boundaries while allowing METR to speak publicly about the rest.

Still, the effectiveness of the model will depend on how often such redactions are requested and whether outside observers see them as reasonable. If redaction requests become too broad, the arrangement could lose much of its value. If they remain targeted, the setup could become a rare example of meaningful AI transparency.

What this could mean for the rest of the AI industry

If Anthropic’s experiment works, other AI developers may be pressured to follow with comparable access for external evaluators. That would move independent oversight from a voluntary virtue signal to an expected operating norm among frontier labs.

It could also reshape how safety claims are made. Instead of relying solely on company-authored model cards, blog posts, or benchmark reports, the industry may need third-party assessments with real inside visibility to remain credible.

For regulators, the arrangement offers a preview of a more structured oversight regime. For investors and enterprise customers, it provides a potential signal that a company is serious about managing model risk. For the public, it may offer a limited but real window into how AI systems are governed before they are broadly deployed.

Bottom line

Anthropic’s decision to give METR employee-level access is a meaningful test case for AI accountability. It does not solve the transparency problem on its own, but it does show one major lab is willing to let outsiders inspect its work far more closely than before.

As frontier AI systems become more capable, that kind of access may be less of an experiment and more of a necessity.