Anthropic gives outside watchdog staff access to its AI safety work

In short: Anthropic has let an outside nonprofit, METR, work inside the company and review safety risks of its AI models with access similar to employees.

What happened

Anthropic, the company behind the Claude chatbot, said it is giving external evaluators “employee-like access” to test and review risks in its AI systems. The evaluator group is METR, a nonprofit that has worked with several major AI labs.

Anthropic says six METR staff have been onboarded. They will have company hardware, office space, and access to internal tools and systems similar to what Anthropic’s own risk teams use. Anthropic also told employees to share information with METR as they would with a coworker.

METR is not being paid by Anthropic. METR also says it does not take donations from AI companies or their staff, and it is funded by individuals, nonprofits, and some government bodies, including the UK’s AI Security Institute.

METR will publish its analysis, and Anthropic cannot block it. Anthropic can ask for specific redactions, like customer data, certain legal communications (attorney-client privilege, which is like confidential talks with your lawyer), or details about how the models are built.

Anthropic compared the setup to bank supervision, where regulators place examiners inside big banks to spot risky behavior early (like having an inspector in the kitchen, not just tasting the food later).

Why it matters

As AI systems get more capable, more people worry about harmful uses, including cyber attacks and other serious misuse. Letting an independent group look at internal systems could make safety checks more credible than company-only reviews. Critics still argue this kind of arrangement can create conflicts of interest, and that stronger government rules may be needed.

Source: Arstechnica