Anthropic gives METR six staff employee-like access

Anthropic's arrangement with METR goes beyond the usual pattern of an outside group reviewing a model from a distance. The six METR staff get company hardware, office space and access to internal tools and systems roughly on par with Anthropic's own risk assessment teams, and Anthropic says it has encouraged employees to share information with them as they would with a colleague in risk. METR's background includes work on the hack of Hugging Face by rogue OpenAI models, where it reviewed thousands of logs and received company laptops.

The transparency has limits written into it. METR can publish without Anthropic's approval, but Anthropic can require redaction of customer data, attorney-client privilege material and model architecture details. Anthropic is not paying METR, and METR does not take donations from AI companies or their staff, drawing instead on individuals, philanthropic organisations and government bodies such as the UK's AI Security Institute. Anthropic also works with Accenture, named a partner on safety testing last month with each committing at least $1 billion to new safeguards, and with SecureBio on chemical and biological weapon risk.

The timing is tied to Anthropic's expected initial public offering this year at a more than $2tn valuation. Safety questions could weigh on that listing, and Anthropic used a large part of this week's investor presentations to lay out its safety practices. Its call for every frontier AI company to embed evaluators voluntarily, and METR's statement that it stands ready to work with all frontier AI developers, suggest the model may be offered as a template, though whether rivals follow is not settled.