Anthropic gives METR internal access to audit AI safety

Anthropic granted six METR researchers internal access to independently review its models and safety practices, with limited redactions.

Anthropic is giving six researchers from nonprofit evaluator METR company hardware, office space, and internal tools to review its AI models, safety practices, and risk management, in access TechCrunch calls employee-level and The Decoder says roughly matches Anthropic's own risk teams. METR will publish its assessments independently. TechCrunch says Anthropic cannot veto publication but may require narrow redactions in predefined categories such as customer data and model architectures, while The Decoder says Anthropic may request redactions and METR can publicly dispute whether they were justified, and also lists attorney-client material as off limits. TechCrunch describes METR as the sole top-tier evaluator in a three-tier system with narrower access for groups such as SecureBio; The Decoder says the deal exceeds access given to SecureBio, Accenture, and Faculty. The move follows a September proposal by CEO Dario Amodei and a White House accord—called the Accord on Super Intelligence by The Decoder—that TechCrunch says was also signed by OpenAI, Google, Meta, Nvidia, and xAI, with OpenAI pledging similar access.

• Six METR researchers receive company hardware or laptops, office space, and internal risk tools, described as employee-level access or roughly equal to Anthropic's own risk teams.
• METR will publish independent assessments of models, safety practices, risks, and incidents.
• Sources differ on redactions: TechCrunch says Anthropic cannot veto publication but may require narrow redactions; The Decoder says Anthropic may request them and METR can publicly say whether they were justified.
• Restricted categories include customer data and model architectures; The Decoder also lists attorney-client or privileged legal material as off limits.
• TechCrunch calls METR the only top tier in a three-tier system, citing SecureBio as a lower tier; The Decoder says access exceeds that given to SecureBio, Accenture, and Faculty.
• The arrangement follows a September proposal by CEO Dario Amodei and a White House outside-review accord that TechCrunch says was also signed by OpenAI, Google, Meta, Nvidia, and xAI, with OpenAI pledging similar access.