Anthropic Grants METR Researchers Employee-Level Access and Waives Veto Rights on Audit Findings
Anthropic has granted researchers at AI evaluation nonprofit METR employee-level internal system access and waived its rights to block independent audit finding
On October 15, 2026, artificial intelligence lab Anthropic officially granted researchers from the non-profit evaluation group METR (Model Evaluation and Threat Research) employee-level access to its internal systems and data. Under this new partnership, Anthropic has also formally waived its right to block, censor, or veto the publication of METR's independent evaluation findings.
Image source: TechCrunch / note.com
Breaking away from the conventional industry practice where frontier AI labs offer third-party auditors only curated documentation or restricted API endpoints, this agreement represents an unprecedented precedent in opening internal corporate infrastructure to an independent safety monitor.
Employee-Level Internal Access and Independent Risk Evaluation
The core breakthrough of this partnership lies in the depth of access granted to external safety researchers.
METR (Model Evaluation and Threat Research) is an independent non-profit research organization specializing in evaluating frontier AI model risks, security and operational incidents, and internal practices. Traditional third-party evaluations have historically operated under structural constraints, relying on curated materials or restricted API endpoints that prevented auditors from observing real-world model behavior.
- Employee-Equivalent System Access: METR researchers receive internal access on par with full-time Anthropic staff, covering internal systems and data.
- Independent Risk Assessment: The external research team evaluates model risks, security and operational incidents, and internal practices entirely on its own terms, free from corporate editorial intervention.
This structural access enables researchers to move past the limitations of selective disclosure and "black-box" auditing, providing the foundation for genuine transparency.
Waiving Publication Veto Rights and Broader Industry Repercussions
Alongside deep system access, the most consequential aspect of the arrangement is Anthropic's voluntary forfeiture of publication veto power.
Historically, even when tech firms permitted external reviews, contracts routinely reserved the company's right to withhold or redact unfavorable findings under the justification of proprietary trade secrets or brand reputation protection. Anthropic has formally relinquished any authority to suppress or alter METR's audit conclusions prior to public release.
According to reporting by TechCrunch, the move is comparable to "handing external inspectors the keys to the corporate kitchen while giving up the power to shut down their reports." This establishes a direct benchmark that puts transparency pressure on competing frontier AI developers and the broader industry to demonstrate comparable openness in independent safety oversight.
Voluntary Corporate Precedent and Operational Security Boundaries
While the agreement sets an important milestone for AI governance, practical limitations and operational caveats remain.
- Voluntary Single-Company Initiative: This arrangement is a voluntary bilateral partnership by Anthropic, not an enforced regulatory mandate or an industry-wide governance standard. Whether other major labs will follow suit remains uncertain.
- Security Boundaries and Trade Secrets: Although Anthropic waived veto rights over evaluation findings, managing intellectual property (IP) protection, commercial secrets, and infrastructure cybersecurity boundaries during real-world audits requires carefully calibrated operational protocols that demand ongoing verification.
The ultimate efficacy of this external auditing model will be judged by the depth and independence of the forthcoming evaluation reports METR publishes under the agreement.