Anthropic grants outside evaluator METR employee-level access in AI safety push

Anthropic, the developer of Claude models, has granted nonprofit METR employee-level access to its systems, signaling an effort to safely advance AI technology without hindering its development. The company is also in talks with other external evaluators, including Hugging Face, SecureBio, and Accenture's AI business Faculty. Anthropic's partnerships with these evaluators involve different levels of access, with METR achieving employee-like access, which includes company hardware, office space, and internal system access, though customer data, legally privileged information, and model architecture remain off-limits.

ANTHRIPC's partnerships with external evaluators aim to scrutinize the technology's risks, despite potential disputes over intellectual property and concerns regarding the independence of evaluators.