Six researchers from METR, a nonprofit that evaluates frontier AI models for catastrophic risk, are now working inside Anthropic with company hardware, office space, and access to the same internal tools as the lab’s own risk assessment teams — and the legal right to publish what they find without Anthropic’s approval. The arrangement, formalized in a post on Anthropic’s website Thursday, arrives one day after a Reuters investigation revealed that the lab had sought strict editorial authority over evaluator reports as recently as October 5.

The timing is not incidental. Anthropic’s blog post establishes a three-tier framework specifying — in writing, for the first time — exactly what embedded evaluators at each level are entitled to see, access, and publish. METR sits at the top tier. What that tier means in practice, and what it still cannot do, is the story of how much the AI safety oversight debate has moved in 10 days.

Anthropic frames the arrangement as a way to close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands — giving independent parties deep access to internal information so they can share evidence and reasoned conclusions about the risks of building frontier models.

What Employee-Like Access Gives METR Researchers

Employee-like access is an unusually specific commitment for an industry whose transparency pledges often lack operational detail. Under the terms Anthropic published Thursday, the six embedded METR researchers receive company hardware, physical office space inside Anthropic facilities, and access to internal tools and systems roughly comparable to what the lab’s own risk assessment teams use.

That access excludes equally specific categories: customer data, attorney-client privileged information, and information about model architectures, unless Anthropic explicitly approves sharing architecture details in a given case. Outside those carved-out categories, Anthropic employees are instructed to share information with the embedded METR researchers as they would with a colleague working on alignment risk.

A firewall provision limits information leakage in the other direction. Anthropic’s engagement with METR runs through a small team operating under enforceable agreements; other METR employees — including researchers who work with Anthropic on separate projects — are siloed from the embedded team’s findings. That separation is designed to ensure the rest of METR’s organization cannot be used as an indirect channel for information the six embedded researchers obtained under confidentiality.

Anthropic’s Three-Tier Evaluator Framework, Explained

Beyond the METR arrangement, Thursday’s announcement formalized a tiered access structure that Anthropic said will govern all its evaluator relationships going forward.

Tier 1: Evaluators conduct narrow, scoped reviews. They may not publish findings without Anthropic’s approval.

Tier 2: Evaluators receive more information and broader access than Tier 1. They can publish without Anthropic’s approval; Anthropic may require redaction of specific pre-defined categories, and evaluators may publicly characterize whether they believe the redactions were justified.

Tier 3: Evaluators receive the full employee-like access described above, in addition to retaining all Tier 2 publication rights.

The publication-rights clause in Tier 2 and Tier 3 is the structural core of what distinguishes this framework from previous industry arrangements. METR can call out redactions it believes are unjustified — as long as it does not reveal the redacted content — creating a mechanism for surfacing disputes even when the underlying information cannot be disclosed. That right is unconditional; Anthropic cannot revoke it, per the published framework.

Two other organizations currently occupy lower tiers. SecureBio, a biosecurity nonprofit, is at Tier 1 and is reviewing the chemical and biological weapons chapter of Anthropic’s August Risk Report, following a completed pilot review of a prior risk report. Accenture’s Faculty business — which Anthropic announced in September as its first embedded evaluator, with each party committing at least $1 billion over five years — is working toward an ambitious partnership that includes red-teaming, alignment assessments, and safeguard testing, but has not yet reached Tier 3 status. Anthropic said it is also in early discussions with Hugging Face.

Anthropic acknowledged that industry-wide evaluator standards do not yet exist and that its own approach is expected to evolve.

The Day Before: Who Wanted Veto Rights?

The framework Anthropic published Thursday has a specific predecessor problem. A Reuters investigation published October 13 found that as recently as October 5 — while publicly pledging editorial independence to evaluators — Anthropic and OpenAI had both privately sought strict editorial authority over what safety evaluators could release.

Anthropic denied that Reuters’ characterization accurately described all its internal discussions, but stopped short of flatly denying the underlying conduct. OpenAI separately acknowledged that it had required the removal of some details from at least one evaluation report — a METR and Redwood Research review of a July Hugging Face security incident in which 700 autonomous AI agents built coordination infrastructure, conducted a multi-day intrusion, and obtained administrator access, according to TechTimes prior reporting.

By October 8, according to the same Reuters account, Anthropic’s internal discussions had shifted direction. Thursday’s published framework is the result.

AI ethics researcher Rumman Chowdhury had warned that labs could use access control as a form of soft leverage — giving evaluators enough access to feel privileged, then implicitly threatening to revoke it if published reports were unfavorable. The Tier 2 and Tier 3 publication provisions are a direct structural response to that critique: Anthropic’s ability to chill publication by threatening to revoke access is limited, at least formally, by the written and enforceable right to publish.

Who Is METR — and Is It Truly Independent?

METR — which stands for Model Evaluation & Threat Research and is pronounced “meter” — is a Berkeley, California-based 501(c)(3) nonprofit that develops methods for assessing whether frontier AI models pose catastrophic risks, then applies those methods to evaluate specific models. It was founded as ARC Evals around 2021 and renamed METR as it expanded its scope. Beth Barnes, who founded the organization, remains CEO.

The organization’s evaluation work focuses on four risk categories: autonomous replication, cybersecurity capabilities, chemical and biological weapons risk, and AI research-and-development acceleration. It is a member of the NIST AI Safety Institute Consortium and provides technical assistance to the European AI Office, per CASRAI’s evaluator independence analysis. Its key evaluation partners include OpenAI, Anthropic, Google DeepMind, and Meta.

In August 2026, METR announced it had raised approximately $71 million in commitments over the previous six months, drawn from philanthropic foundations and individual donors rather than the AI companies whose models it evaluates. That funding structure is METR’s core answer to the auditor-pays problem — the well-documented dynamic in financial auditing in which auditors paid by the companies they audit have structural incentives to produce favorable opinions.

METR’s stated policy is that it does not accept compensation from the AI companies whose models it evaluates — a commitment documented in its publicly available conflict-of-interest policy.

The independence critique, however, is not primarily about direct financial relationships. Former White House AI advisor David Sacks has publicly argued that METR is not genuinely independent of Anthropic because of the dense web of relationships connecting the two organizations through the effective altruism community — the philosophical and philanthropic movement from which both drew significant founding personnel and early funding support. Critics have pointed to the role of Good Ventures and Open Philanthropy as major funders across the AI safety nonprofit world, and to Anthropic co-founder Dustin Moskovitz’s history of supporting similar organizations, as evidence that the evaluator and the evaluated are embedded in the same institutional network regardless of direct financial separation, according to prior TechTimes coverage of the debate.

Cybersecurity researcher Joshua Saxe disputed that framing, warning that “vilifying METR is not a step towards making the world a better place” and calling it “one of the few nonprofit centers of AI safety excellence” whose research rigor should be the relevant measure of its independence.

Over 200 researchers and practitioners — including Turing Award winner Geoffrey Hinton — have signed an open letter from the AI Evaluator Forum calling for independence standards that address both financial relationships and network ties. The letter predates Thursday’s announcement; whether Anthropic’s published framework meets its criteria will likely be a subject of near-term debate.

What METR Can Publish — and What It Cannot Force

Thursday’s framework resolves one question clearly and leaves another open.

The resolved question is editorial: METR has the unconditional right to publish its findings, to characterize any redactions Anthropic requires, and to state publicly whether it believes those redactions were justified. No prior external safety evaluation of a frontier AI lab has included that right in a binding form.

The unresolved question is enforcement. METR can investigate and publish; it cannot compel Anthropic to halt model training, delay a release, or take any specific remedial action. Anthropic has stated explicitly that the goal is not to transfer accountability — the lab remains responsible for the safety of its models — and that independent evaluators add value by identifying blind spots and verifying whether the company is following its own stated safety commitments, not by acquiring authority over what the company does next.

That distinction places the Anthropic model closer to the external audit tradition than to the embedded bank examiner model Dario Amodei has cited as his template. The Federal Reserve’s resident examiners at systemically important banks have statutory enforcement authority — they can issue mandatory findings and require remediation. METR has disclosure authority only. The enforcement gap is the dimension the legislative process is now working to fill.

The FRONTIER Act (HR 9925), introduced July 23, 2026 by Representatives Obernolte and Trahan, would mandate third-party audits for frontier AI models at the federal level. A separate bill from Senators Thune, Cruz, and Klobuchar would give the government authority to block unsafe model releases if a company fails its duty of care. Both bills remain in committee. Anthropic’s three-tier framework is, as of Thursday, the only published industry specification of what evaluator access and publication rights should look like — which means it is also the model legislators would be codifying, or departing from, if either bill advances.

Anthropic has noted it is a signatory to the White House Accord on Superintelligence, under which participating companies pledged to partner with independent external auditors. The Accord carries no enforcement mechanism. Thursday’s framework is the first binding specification of what that pledge means at the operational level — for Anthropic and METR, at least.

Whether OpenAI, Google DeepMind, and Meta publish equivalent frameworks, or whether the Anthropic model becomes a differentiating standard rather than an industry floor, will determine how much Thursday’s announcement reshapes the governance landscape it is now part of.

Frequently Asked Questions

What is a Tier 3 evaluator under Anthropic’s new framework?

A Tier 3 evaluator — the highest level in Anthropic’s published three-tier structure — receives employee-like physical and systems access to the lab, including company hardware, office space, and access to internal tools comparable to what Anthropic’s own risk assessment teams use. Tier 3 evaluators also retain the publication rights established at Tier 2: they can publish findings without Anthropic’s approval, and they can publicly characterize any redactions Anthropic requires, including stating whether they believe those redactions were justified. METR is currently the only organization at Tier 3.

Is METR truly independent of Anthropic?

METR is a 501(c)(3) nonprofit that states it does not accept funding from the AI companies whose models it evaluates, addressing the direct financial independence question. The more contested independence question is network-based: METR emerged from the effective altruism research community that also produced many of Anthropic’s founders, and shared funders — including Good Ventures and Open Philanthropy — are major backers across the AI safety nonprofit world. Critics including former White House AI advisor David Sacks argue that those ties constitute a structural dependence that financial separation alone cannot resolve. Supporters like cybersecurity researcher Joshua Saxe argue the relevant measure is the rigor of METR’s published research, not its personnel network. The AI Evaluator Forum’s open letter — signed by over 200 researchers including Geoffrey Hinton — called for formal independence standards that would address both dimensions.

Can METR force Anthropic to halt training or delay a model release?

No. METR’s authority under the published framework is disclosure authority, not enforcement authority. It can investigate, publish findings, and characterize redactions — but it cannot compel Anthropic to take any specific action in response to those findings. Anthropic retains full operational control over its models. The enforcement gap is what the FRONTIER Act and the Senate duty-of-care bill are designed to fill at the federal level, though both bills remain in committee.

What makes today’s announcement different from past AI safety pledges?

Most previous AI safety commitments were statements of intent without operational specifications — pledges to work with evaluators without defining what evaluators could access or publish. Thursday’s framework is, as of publication, the only written industry specification of what evaluator access tiers look like in practice and what publication rights each tier carries. Critically, it includes an explicit prohibition on Anthropic vetoing METR’s publications — a right the lab had reportedly sought to retain as recently as October 5, according to Reuters. Whether other frontier labs publish equivalent frameworks, or whether this remains a single-lab standard, will determine how significant that distinction proves.