Back to the stories

Anthropic gives METR employee-like access to evaluate its AI systems and safety practices

Score 10.0,

AI safety research

For the first time, an external nonprofit will be given employee-like access. That lets them inspect a frontier AI lab from the inside.

On October 18, 2026 Anthropic announced it is granting METR employee-like access. The role is to evaluate the safety of its systems.

The announcement also created the Anthropic Public Benefit Fund. Anthropic said it is in early talks with other groups, including Hugging Face, about similar access.

This is a governance move, not a capability release. It changes who can verify what the lab is doing.

Safety checks can now include independent inspections of internal practice, not just tests the company publishes for the public.

Think of it as giving a trusted inspector a temporary employee badge. METR can review internal tests and logs. They can examine development practices rather than only model outputs.

That does not mean the models are now public or that anyone can run them. The access is limited and negotiated, and it is practical only for organisations with the expertise and legal agreements to handle sensitive material.

For regulators, journalists and rivals, this raises the bar. Credible independent evaluation can happen without relying on a company to publish selective tests.

The question is whether other labs will follow. Will these arrangements become routine audited processes or stay as ad hoc handshakes? Watch for more groups taking similar access, formal benchmarking of what inspectors find, and any regulatory attention.

The real test will be whether independent teams can turn inspection rights into publicly trusted findings.