Anthropic Gives METR Employee-Level Access for AI Safety Reviews

Introduction

Anthropic is moving beyond broad promises about external oversight and giving an independent evaluator a much closer look at its operations. The nonprofit METR has placed six staff members inside Anthropic’s office, where they receive company hardware, workspace, and access to internal tools and systems broadly comparable to those used by Anthropic’s own risk-assessment teams. The arrangement makes the evaluators more like embedded safety colleagues than occasional outside consultants.

Key points

• Access is designed to be employee-like. METR’s team can use internal systems and is encouraged to receive information from employees in the same way an internal risk group would.
• METR retains publication rights. The evaluator will be able to publish its analysis, and Anthropic cannot veto publication. Anthropic may request redactions involving customer information, attorney-client privilege, or model architecture details.
• Anthropic is not funding the arrangement. METR says it does not accept donations from AI companies or their employees. Its support comes from individuals, philanthropic organizations, and government bodies, including the UK’s AI Security Institute.
• The broader industry has not matched the commitment yet. Anthropic’s chief executive has advocated embedding outside auditors in frontier labs. OpenAI has voiced support for the idea but has not announced an equivalent implementation plan.

Why it matters

More capable models may create risks involving cybersecurity, the economy, and other areas of public safety. At the same time, laboratories cannot necessarily disclose every detail of their testing: some information could help malicious actors, reveal confidential technical designs, or expose customer data. Anthropic compares its approach with resident supervision in banking, where outside examiners work inside major institutions to identify dangerous risk-taking before it becomes a wider problem.

The important question is not simply whether an outside group is present, but what it can actually see and how freely it can report what it finds. METR has previously worked with major AI developers and participated in investigations of unusual model behavior, including a review of logs and company devices in the Hugging Face incident involving rogue OpenAI models. That experience may help it handle sensitive evidence, but embedded access also creates difficult questions about independence and institutional pressure.

Anthropic’s model establishes a middle ground between complete secrecy and unrestricted disclosure. The company says it wants to be publicly accountable, while reserving limited redaction rights. Critics may still argue that voluntary arrangements cannot replace government oversight, particularly if evaluations concern systems with effects beyond a single company.

The company is also working with Accenture on safety testing and with SecureBio on chemical and biological weapon risks. As Anthropic seeks investor confidence ahead of an expected public offering, safety governance has become a business and reputational issue as well as a technical one. Whether this experiment becomes a model for other frontier labs will depend on the substance of METR’s reports, the quality of disclosure, and the eventual role of formal regulation.

Source: Ars Technica AI