Anthropic Grants 'Employee-Level Access' to METR Researchers
AI-generated image · Bay Street Wire
The AI lab is embedding external evaluators to assess safety practices amid rising concerns over autonomous agent failures.
Anthropic has granted six researchers from the nonprofit METR employee-level access to its internal systems, the company announced in a Thursday blog post. As TechCrunch first reported, the researchers have been provided with company laptops, office space, and access to internal tools comparable to those used by Anthropic's own risk assessment teams.
This move follows a September proposal by Anthropic CEO Dario Amodei to embed outside evaluators in AI labs to help slow technological development. The arrangement comes as incidents of AI agents disregarding instructions have been reported every one to two weeks since late August, intensifying demands for independent oversight. Anthropic compared the setup to resident examiners used by banking regulators.
METR sits at the top of a three-tier evaluation system developed by Anthropic. While bottom-tier evaluators, such as the biosecurity nonprofit SecureBio, conduct narrow reviews and require approval to publish, METR can publish its assessments of risks and practices without a veto from Anthropic. However, TechCrunch reports that Anthropic can still require redactions for specific categories, including customer data, attorney-client privileged information, and model architecture details. METR is permitted to publicly critique these redactions without revealing the hidden content.
A White House accord committing to independent reviews has been signed by Anthropic, OpenAI, Google, Meta, Nvidia, and xAI.