Anthropic has installed a team of outside evaluators inside its offices, giving the nonprofit METR access to company hardware and internal systems as part of an effort to have independent reviewers assess the risks posed by its Claude AI models. The arrangement, announced Thursday, marks the first concrete step toward implementing a proposal Anthropic chief executive Dario Amodei made in September to embed external auditors at leading AI labs. METR has previously worked with OpenAI, Anthropic, Google and Meta.
Six METR employees have been onboarded at Anthropic and given company hardware, office space and access to internal tools and systems that the company says are roughly equivalent to what its own risk assessment teams use. Anthropic said it has encouraged employees to share information with the evaluators as they would with a colleague working on risk. The lab is not paying METR for the arrangement, and the research group does not accept donations from AI companies or their staff, relying instead on individual and philanthropic donors as well as government bodies such as the UK's AI Security Institute.
Under the terms, the evaluators will publish their analysis and Anthropic cannot veto publication, though the company can require redaction of certain information such as customer data, material covered by attorney-client privilege or details about model architectures. Anthropic said it wants to be publicly accountable for its decisions but cannot disclose all information relevant to its safety work, arguing that public disclosure could aid malicious actors or competitors. The company compared the setup to bank supervision, in which regulators assign resident examiners to work inside the largest banks.
Amodei's proposal drew rare agreement in September from OpenAI chief executive Sam Altman and SpaceX chief executive Elon Musk, who both backed the idea of embedding external auditors in leading labs. OpenAI has not yet outlined plans to bring in evaluators, according to the report. Critics of the approach have argued that such arrangements could create conflicts of interest, pointing to problems in the accounting sector, and that stronger government regulation is needed instead.
The arrangement comes as Anthropic faces pressure to reassure investors that safety issues will not overshadow its initial public offering, which is expected this year at a valuation of more than $2tn. The company devoted a large portion of its investor presentations this week to its safety practices. METR was among the organizations that investigated the hack of Hugging Face by rogue OpenAI models, reviewing thousands of logs and receiving company laptops, and Anthropic said it was still working with Accenture, named a safety testing partner last month, with each committing at least $1 billion to new safeguards. Another third-party tester, SecureBio, reviews chemical and biological weapon risk for the company.
Anthropic said it is embedding the evaluators voluntarily and believes every frontier AI company should do the same. METR said it stands ready to work with all frontier AI developers to bring better evidence to the public.
More AI news from TechManNews.
