Anthropic Gives Six METR Staff Internal Access for Safety Reviews

Anthropic has begun giving six METR employees access to company equipment, office space and some internal systems, opening parts of its safety work to outside review.

The access level is designed to resemble that of Anthropic’s risk assessment teams. Anthropic said Thursday, Oct. 15, that it had begun giving six METR employees access, but did not give a start date. The arrangement lets METR examine internal information beyond tests of publicly available models.

Customer data and attorney-client privileged information are excluded. Model architecture information is restricted, though Anthropic may share it when highly relevant after internal approval. Employees are encouraged to share other information with METR as they would with colleagues working on alignment risks.

METR will publish analysis of Anthropic’s risks, incidents and safety practices. Anthropic cannot veto publication, but it can require METR to redact information in narrowly defined categories. METR can publicly explain whether it considers any requested redactions justified.

The arrangement is voluntary and does not constitute a comprehensive regulatory audit, a safety certification or approval for a model release. Anthropic’s Oct. 15 announcement described access as underway and did not report a completed assessment.

METR has cautioned that an outside review depends on access granted by the company and may not reach the same information available to Anthropic’s senior safety staff. Access limits and redactions could affect the evidence available to the public.

METR, a nonprofit that evaluates frontier AI capabilities and risks, previously reviewed Anthropic risk reports and red-teamed an internal agent-monitoring system. Separately, METR has proposed public reviews of frontier AI loss-of-control risks.

Anthropic also works with SecureBio and Accenture and has held preliminary discussions with Hugging Face. Those collaborations have different access and publication terms.