Anthropic Grants AI Watchdog Employee-Level Access

Anthropic has handed a team from METR, a nonprofit AI safety evaluator, what it calls "employee-like access" to its offices, hardware and internal systems, a step beyond the usual arm's-length audits that outside groups typically get. METR has already worked with OpenAI, Google and Meta, but the embedded arrangement with Anthropic goes further than prior collaborations, giving outside reviewers a seat inside the building rather than a data dump to pore over remotely.

The move follows a September moment when Anthropic chief Dario Amodei found rare agreement with Sam Altman and Elon Musk on the need for third-party auditors inside frontier AI labs. Thursday's announcement is the first concrete follow-through on that pledge. OpenAI, despite Altman's public support for the idea, hasn't yet detailed its own plans to bring in similar evaluators.

The push comes as labs face mounting pressure to show they can catch dangerous capabilities before models are released, particularly as systems grow more capable of acting with less human oversight. Anthropic's answer is to let outsiders in, but the company has also acknowledged limits on what it can disclose, setting up tension between transparency and the secrecy that still surrounds frontier model development.