Anthropic has placed a team from the nonprofit METR inside its headquarters, granting the external watchdog access to the company’s hardware, code repositories and internal monitoring systems. The arrangement gives METR “employee-like” privileges, allowing it to examine model behavior and risk controls directly. Unlike typical audits, METR may publish its analysis without Anthropic’s prior approval, subject only to limited redactions.
The move follows a series of industry pledges to expand third-party evaluation after growing alarm over the economic, cybersecurity and existential risks posed by increasingly capable AI. Anthropic’s chief executive Dario Amodei secured backing from OpenAI’s Sam Altman and SpaceX founder Elon Musk for a proposal that would embed external auditors in leading labs. OpenAI has not yet detailed how it will implement similar oversight.
METR, a nonprofit that has previously collaborated with OpenAI, Google, Meta and other leading labs, was among the groups that investigated the recent Hugging Face breach involving rogue OpenAI models, reviewing thousands of logs and receiving company laptops for analysis. The organization does not receive funding from AI firms or their staff; instead it relies on donations from individuals, philanthropic foundations and government bodies such as the United Kingdom’s AI Security Institute. Anthropic does not pay METR for the embedded role.
Anthropic argues that public disclosure of internal details could aid malicious actors or competitors, and therefore limits redactions to categories such as customer data, attorney-client privileged material and proprietary model architecture. The company likens the arrangement to bank supervision, where regulators station resident examiners inside major financial institutions to detect risky behavior before it threatens the public. Employees have been encouraged to treat METR auditors as internal risk-team members, sharing information accordingly.
Critics warn that embedding external reviewers could create conflicts of interest similar to those observed in the accounting sector, and they call for stronger governmental oversight rather than voluntary arrangements. They contend that the close proximity of auditors to the organizations they evaluate may compromise independence, especially when the same firms are also pursuing large safety-related investments with partners such as Accenture and SecureBio.
Anthropic is simultaneously preparing for an initial public offering projected for this year at a valuation exceeding $2 trillion, and it has highlighted safety practices in recent investor briefings. The company continues to work with Accenture, which was named a safety-testing partner, with each party committing at least $1 billion toward new safeguards. SecureBio also serves as a third-party tester focused on chemical and biological weapon risk, reinforcing Anthropic’s broader strategy to embed multiple independent evaluators.