Anthropic gives outside evaluators employee-like access
What changed
Anthropic says a team from METR, a nonprofit that tests frontier AI models, now has access close to what its own staff have. METR will publish what it finds. Anthropic cannot stop that, but it can ask METR to remove some kinds of information. It also works with SecureBio and Accenture, and is talking to other groups.
Why it matters for builders
If you rely on an AI provider, outside tests are new evidence to read when you decide how far to trust a model. The company says evaluators need four things: real skill, independence, enough access, and the freedom to publish. That is a good checklist for your own systems too.
What to check
This is a company's own announcement of its plans. It describes access, not results. Wait for the published findings before you decide anything about safety.
Go deeper
- Autonomy is earned: how much freedom should an AI agent get?
- 50 tests passed. Is it safe? Why no failures is not proof