Anthropic gives nonprofit evaluator METR employee-level access to its systems

Centre · 2Anthropic said in a blog post on Thursday 15 October that it has given six employees of METR (Model Evaluation and Threat Research), a nonprofit that tests frontier AI models, employee-level access to its systems so they can independently assess its models, safety practices and risk management. 1 2 Centre · 2The METR team has Anthropic-issued laptops, office space and internal tools roughly on a par with those used by Anthropic's own risk assessment teams, and Anthropic employees are encouraged to share information with them as they would with a colleague working on alignment risk. 1 3 Centre · 2Anthropic said it cannot veto METR's findings on its risks, incidents and practices, but can require narrowly defined redactions, such as customer data, legally privileged information and model architecture details; METR may publicly question those redactions as long as it does not reveal the withheld material. 3 1

Centre · 2Anthropic is developing three tiers of access for outside evaluators: the first for narrowly scoped investigations that need approval to publish, the second with more information and publication without approval but subject to redactions that evaluators can challenge publicly, and the third, which only METR occupies, with employee-like access and publication rights similar to the second tier. 3 1 Centre · 1Biosecurity nonprofit SecureBio is in the first tier and is reviewing the chemical and biological weapons chapter of Anthropic's August risk report, and Anthropic said it would consider widening its access and publication rights afterwards; Anthropic is also working with Accenture's AI business Faculty and is in talks with others including Hugging Face, and described those partnerships as early-stage with more limited access. 3 Centre · 1Anthropic compared the arrangement to banking, where regulators place resident examiners inside the largest banks, and said it wants to be "publicly accountable for the decisions we’re making". 1

Centre · 1TechCrunch said the announcement fulfils a promise CEO Dario Amodei made in a September essay proposing outside evaluators embedded in AI labs, that OpenAI quickly pledged the same, and that both companies, along with Google, Meta, Nvidia and xAI, have signed a White House accord committing to independent review; it also noted that incidents of AI agents acting outside their instructions have been disclosed every week or two since late August. 1 Centre · 1CNA, drawing on Reuters reporting, put the move in the context of a dispute over control: Reuters reported earlier in the week that Anthropic and OpenAI had sought significant control over what outside reviewers could publish, fearing leaks of trade secrets, and a source said talks with Anthropic had recently become more open. 3 Centre · 1Open concerns remain: a source told Reuters that information labs consider intellectual property can also matter for assessing risk, leading to disputes and redactions, and experts say evaluators' independence is questionable because Anthropic and OpenAI choose who reviews them. METR did not immediately respond to a request for comment, having proposed a day earlier that developers allow private staff interviews, timely publication of assessments and transparency about access limits and redactions. 3