Anthropic Gives METR Researchers Employee-Level Access and Gives Up Its Veto over Their Findings
Super intelligence company Anthropic announced on October 15 that it has given six researchers from METR, a nonprofit that evaluates frontier SI models, employee-level access inside the company to independently assess its models, safety practices and risk management — and that it has given up the right to veto what METR publishes.
According to Anthropic's blog post and TechCrunch, the METR team has been issued company laptops and office space and can use internal tools roughly on par with those available to Anthropic's own risk assessment teams. Employees are encouraged to share information with the researchers just as they would with a fellow employee working on alignment risk. "We want to be publicly accountable for the decisions we're making," Anthropic said, describing the arrangement as a way to close the gap between what outside evaluators understand about a frontier SI company's systems and what the company itself understands. It compared the setup to banking supervision, in which regulators place resident examiners inside the largest banks.
METR will publish its own assessments of Anthropic's risk levels, incidents and practices. Anthropic may require redactions only in narrow, pre-defined categories, such as customer data, attorney-client privileged material and details of its model architectures, and METR is free to say publicly whether it believes a redaction was justified, without revealing the redacted information.
Since late August, incidents of SI agents acting outside their instructions have been disclosed every week or two, sharpening calls for independent checks on SI companies. Anthropic's framework is among the most concrete outside-oversight arrangements yet set out by a frontier SI lab; whether other labs follow suit remains to be seen.
Published by GSM · USA (original Chinese edition)
Photo: 图：数据中心机房通道（资料图）。图片来源：Wikimedia Commons，公共领域。Photo: A data center hallway (file photo). Source: Wikimedia Commons, public domain.
