Anthropic offers employee‑level access to METR, pledging collaboration with evaluators to review its models and safety practices
ai-safety anthropic
| Source: Techmeme | Original article
Anthropic is granting evaluator-like access to its METR system as part of a pledge to enable independent third‑party assessment of its models and safety practices.
Anthropic announced on October 15 that it has granted a team from METR – a nonprofit that evaluates frontier‑AI models – employee‑level access to its internal tools, hardware and office space. Six METR staff members can now work alongside Anthropic engineers, run the company’s latest models and inspect safety‑related processes. The move follows Anthropic’s pledge made the previous month to “strengthen independent, third‑party assessment of our models and safety practices.”
The decision is significant because it offers one of the few avenues for external auditors to examine a leading foundation model developer’s inner workings in depth. By handing over the same privileges it gives its own employees, Anthropic aims to demonstrate confidence in its safety regime and to address mounting regulatory and public pressure for transparent oversight. The United States Federal Trade Commission’s recent investigative demands on Anthropic and other labs underscore the heightened scrutiny the sector faces.
However, the partnership is not without questions. A September 2026 analysis highlighted concerns about METR’s funding sources and donor ties, suggesting that the nonprofit’s independence could be scrutinized as its assessments are published. Observers will be watching how METR’s risk reports are framed and whether they reveal any material safety gaps.
Going forward, the AI community will track the forthcoming METR assessment results, any subsequent policy adjustments by Anthropic, and whether other frontier‑AI firms adopt similar “employee‑like” access arrangements. The development may also influence ongoing regulatory dialogues in Europe and the United States, where calls for robust third‑party evaluation of powerful models are intensifying.
Sources
Back to AIPULSEN