Anthropic Grants METR Employee-Level Access for Frontier AI Safety
Six METR researchers receive internal access and hardware inside Anthropic under a new three-tier framework governing third-party model evaluation.
In short
- Anthropic onboarded six METR researchers with company laptops, office space, and internal tool access comparable to its in-house risk teams.
- METR retains the independent right to publish its risk findings and characterize whether Anthropic redactions are justified.
- The partnership introduces a three-tier evaluator framework, with METR occupying the highest access level alongside early-stage reviews with SecureBio and Accenture.
- METR holds disclosure powers rather than enforcement authority, meaning it cannot mandate model pauses or training halts.
On October 15, 2026, Anthropic announced that it has granted employee-like access to six researchers from METR, an independent nonprofit evaluating frontier artificial intelligence systems. Under the agreement, the embedded researchers receive company hardware, dedicated office space, and access to internal tooling on par with Anthropic's own risk assessment teams. The initiative makes good on voluntary commitments made under the White House Accord on Super Intelligence to open frontier development to third-party scrutiny.
Anthropic stated that the arrangement aims to close the information asymmetry between external safety evaluators and internal model builders. The move follows growing government scrutiny, including subpoenas issued to Anthropic and OpenAI by the Federal Trade Commission, alongside incident disclosures where autonomous agents operated outside their intended instructions.
Inside the Employee-Level Access Agreement
Under the terms outlined by Anthropic, the six METR staff members operate under enforceable agreements. Anthropic employees are instructed to share information with them just as they would with internal colleagues working on alignment risk. However, strict firewalls remain in place. Information concerning customer data, attorney-client privileged matters, and proprietary model architectures is excluded from baseline access unless explicitly authorized.
To prevent wider leaks, Anthropic established an internal barrier separating the embedded METR squad from the remainder of the nonprofit's staff. Other METR researchers, including those working on separate Anthropic collaborations, are prohibited from receiving data gathered through the embedded channel.
ZDNET noted that the initiative implements a proposal outlined in September 2026 by Anthropic chief executive Dario Amodei, who recommended placing outside evaluators inside labs as a mechanism for pacing frontier development. Anthropic drew a direct comparison to financial supervision, noting that bank regulators maintain resident examiners inside systemic institutions to monitor internal risk profiles that outside observers cannot see.
The Three-Tier Evaluator Framework
Alongside the METR deployment, Anthropic formalized a three-tier access structure designed to standardize its interactions with external auditing groups, as detailed by Tech Times.
In Tier 1, outside organizations execute tightly scoped investigations into specific hazards. Labs retain editorial discretion, meaning evaluators cannot publish without Anthropic's explicit sign-off. Biosecurity research group SecureBio currently operates at this level, reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report following an initial pilot study.
Tier 2 provides broader operational visibility and grants evaluators independent publication rights. While Anthropic retains the right to require redactions for sensitive categories, evaluators can publicly comment on whether they consider those redactions justified. The Decoder reported that Anthropic is working toward evaluative partnerships with Accenture and Faculty, while holding early discussions with Hugging Face.
Tier 3 combines the independent reporting permissions of Tier 2 with the comprehensive physical and operational systems access granted to METR. Anthropic cannot veto METR's findings, though it can mandate redactions for predetermined sensitive areas. In turn, METR maintains the unilateral right to disclose the context of those redactions and evaluate whether they materially obscure the report's conclusions.
Independence and Enforcement Boundaries
METR, founded by chief executive Beth Barnes as ARC Evals before expanding, focuses its technical evaluations on catastrophic hazards such as autonomous replication, cybersecurity breaches, biological weapons capabilities, and accelerated machine learning research. In a policy statement published on October 14, 2026, METR outlined its operational requirements for public assurance engagements, stressing process transparency, independent reporting, and the ability to disclose any institutional pressure encountered during an assessment.
According to Tech Times, METR raised approximately $71 million in philanthropic commitments over six months in 2026 and maintains a formal policy barring direct compensation or donations from the frontier AI developers it audits. Despite these measures, some observers, such as former White House AI advisor David Sacks, have questioned METR's independence, citing overlapping personnel and institutional ties within the effective altruism funding network. Conversely, researchers including Joshua Saxe have defended METR's track record, pointing to its published methodologies as the proper benchmark for objectivity.
A fundamental boundary of the agreement centers on operational authority. While METR has the legal clearance to publish risk analyses without Anthropic's consent, it holds no regulatory or legal power to pause model development, delay deployments, or enforce remedial engineering. As ZDNET reported, the arrangement remains an industry-led voluntary initiative.
Shifting Regulatory Demands
The formalization of Tier 3 access comes against a backdrop of increasing legislative and regulatory movement in Washington. The FRONTIER Act, introduced in the U.S. House of Representatives in July 2026 by Representatives Obernolte and Trahan, seeks to mandate independent verification organizations to examine frontier development workflows. In the Senate, legislation sponsored by Senators Thune, Cruz, and Klobuchar would give the government authority to block unsafe model releases if a company fails its duty of care.
At the same time, the federal government's Super Intelligence Force recently announced requirements for developers to report safety incidents swiftly following a public incident in which Claude generated an inaccurate homicide tip. Because existing executive accords lack enforcement teeth, Anthropic's tiered structure establishes the first operational template for how embedded external evaluation might function as formal regulatory standards take shape.
Questions people ask
What specific access do METR researchers receive inside Anthropic?
Six METR researchers have been provided Anthropic-issued laptops, dedicated physical office space, and access to internal tooling comparable to Anthropic's internal risk assessment teams. Information covering customer data, attorney-client privilege, and core model architectures remains off-limits.
Can Anthropic block METR from publishing its risk assessments?
No. Under the Tier 3 framework, Anthropic cannot veto publication. It can mandate redactions for pre-defined sensitive categories, but METR retains the contractual right to state publicly whether it believes those redactions were justified.
Can METR halt model deployment if it discovers critical risks?
No. METR possesses disclosure authority rather than enforcement authority. It can publish findings and evaluate risk levels, but it lacks the statutory power to halt training runs, block product launches, or compel operational remedies.
How does Anthropic's three-tier framework classify evaluators?
Tier 1 involves narrow reviews with company publication approval required (e.g., SecureBio). Tier 2 grants independent publication rights with redaction terms. Tier 3 provides employee-like system access alongside independent publication rights, with METR currently as the sole participant.
Sources
- An update on independent evaluatorsAnthropic, Oct 15, 2026, official source
- Anthropic gives third-party auditors ’employee-like access’ in AI industry firstZDNET, Oct 16, 2026
- Anthropic is letting METR researchers look under the hood with employee-level accessTechCrunch, Oct 15, 2026
- Anthropic gives METR inside access to audit its AI models and safety practicesThe Decoder, Oct 16, 2026
- Our approach to working with AI companiesmetr.org, Oct 14, 2026
- Anthropic says it’s giving outsiders ‘employee-like access’ in safety pushThe Independent, Oct 16, 2026
- METR Gets Employee-Like Access Inside Anthropic, With Unconditional Right to Publish FindingsTech Times, Oct 16, 2026
How this story was made: the newsroom picked it up from Techmeme, techcrunch.com and the-decoder.com, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (46 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 17, 2026 at 01:27 UTC
