Anthropic has granted the METR nonprofit “employee-like access” to its internal systems and hardware to assess the risks of its artificial intelligence models. This move fulfills a commitment to increase safeguards amid growing concerns that increasingly powerful AI systems could pose catastrophic threats to the economy and public safety.
The arrangement marks the first concrete step in implementing pledges made by leading AI labs to embed external auditors within their operations. Anthropic chief executive Dario Amodei proposed this framework in September, a plan that drew rare support from rivals Sam Altman of OpenAI and Elon Musk of SpaceX. While OpenAI has yet to detail its own plans for external evaluators, Anthropic is moving forward with the initiative.
Deep Integration with METR
METR, a nonprofit research group that has previously collaborated with major technology firms including Google and Meta, now has a team embedded directly in Anthropic’s office. The organization has onboarded six employees who are provided with company hardware, office space, and access to internal tools. Anthropic stated that this level of access is roughly equivalent to what its own internal risk assessment teams possess.
The company has encouraged its staff to share information with the METR evaluators as they would with a fellow employee working in risk management. This deep integration is designed to allow for more rigorous and realistic testing of the Claude model series. METR previously investigated a security incident involving rogue OpenAI models on Hugging Face, where it reviewed thousands of logs and received company laptops for analysis.
Independent Publication Rights
A critical component of this agreement is the independence of the watchdog. METR can publish its findings without Anthropic’s approval. The company cannot veto the release of the analysis, although it retains the right to require redaction of specific sensitive information. This includes customer data, details protected by attorney-client privilege, and proprietary information regarding model architectures.
Anthropic argued that sharing all safety-related information publicly could benefit malicious actors or competitors. The company compared its approach to bank supervision, where regulators assign resident examiners to work inside the largest financial institutions to detect risks in advance. “If a bank’s risk-taking could threaten the public, third parties try to see that in advance,” the company said in a statement.
Critics of the proposal have raised concerns about potential conflicts of interest, drawing parallels to issues found in the accounting sector. Some argue that greater government regulation is necessary rather than relying on voluntary industry measures. Anthropic maintains that voluntary embedding of evaluators is the most effective path forward and believes every frontier AI company should adopt similar practices.
Financial and Regulatory Context
The push for enhanced safety measures comes as Anthropic prepares for its initial public offering, which is expected later this year with a valuation exceeding $2 trillion. The company has devoted significant portions of its recent investor presentations to outlining its safety practices to reassure stakeholders that technical risks will not overshadow its commercial ambitions.
Anthropic is not paying METR for this arrangement. The research group does not accept donations from AI companies or their staff. It is funded by individual donors, philanthropic organizations, and government bodies such as the UK’s AI Security Institute. The company also continues to work with Accenture, which was named a partner for safety testing last month, with both parties committing at least $1 billion to establish new safeguards. Another third-party tester, SecureBio, reviews chemical and biological weapon risks for the company.
“As models have become more powerful, they’ve required stronger safeguards and more challenging risk assessment,” Anthropic said. “We want to be publicly accountable for the decisions we’re making. But we can’t publicly share all the information relevant to our safety work.” METR added that it stands ready to work with all frontier AI developers to bring better evidence to the public.
Source: Ars Technica
