As artificial intelligence models grow increasingly advanced, questions surrounding their safety and potential societal risks have become more difficult to address. In response, AI developer Anthropic has announced it will provide a research nonprofit organization known as METR with unprecedented access to its internal systems to scientifically measure potential catastrophic threats.
This initiative follows an essay by Anthropic CEO Dario Amodei, in which he suggested that major AI laboratories should permit third-party supervisors, or embedded evaluators, to monitor their operations. Subsequently, industry leaders and government officials signed a voluntary agreement at the White House emphasizing industry self-policing and independent external audits to verify that safety controls function as intended.
Anthropic has now onboarded six staff members from METR, granting them operational tools, systems, and information comparable to what the company's own risk assessment teams utilize. Crucially, Anthropic stated that METR will retain the ability to publish evaluations regarding company risks, incidents, and safety practices without requiring clearance from Anthropic itself.
AI developers typically maintain strict confidentiality over model training details, safety mechanics, and internal risks to prevent malicious exploitation and protect trade secrets. Anthropic noted that granting selected external experts deeper access offers a way to bridge this transparency gap and provide evidence-based conclusions to the public.
Despite the high level of access, certain boundaries remain in place. METR is restricted from viewing customer data, legally privileged files, and specific proprietary details regarding Anthropic's model architectures. While Anthropic holds the right to request the removal of such sensitive material from public reports, METR retains the freedom to publicly disclose whether it believes those redactions were justified. Anthropic emphasized that the arrangement is strictly limited to a small team operating under enforceable legal agreements.
Looking ahead, Anthropic is structuring its independent evaluation framework into three distinct tiers. The first tier involves limited investigations where the company controls data access and approves publications. The second tier grants broader information access with independent publication rights subject to specific redactions. The highest tier reflects the current arrangement with METR, combining near-total internal access with autonomous publishing rights.
In addition to the METR partnership, Anthropic is collaborating with SecureBio to review the chemical and biological weapons sections of its upcoming risk reports. The company is also building evaluation partnerships with Accenture and Faculty, alongside preliminary discussions with other organizations including Hugging Face.