Anthropic gives a six-person METR team employee-like inside access and lets METR publish its risk findings without an Anthropic veto

Why it matters

【Key points】

On October 15 Anthropic said it has given a team from the nonprofit METR "employee-like" access, delivering on last month's pledge of stronger third-party assessment. Six METR staff have company hardware, office space and access to internal tools and systems roughly on par with Anthropic's own risk-assessment teams. Outside a few categories such as customer data, privileged legal material and model architecture, staff are encouraged to share with them as with a colleague.

METR will publish analyses of Anthropic's risk levels, incidents and practices. Anthropic cannot veto publication; it can only require redactions in narrow predefined categories, and METR may say publicly whether it thinks redactions were justified. It set out three tiers: tier one, company-chosen access with approval to publish (SecureBio, reviewing the chem-bio chapter of the August Risk Report); tier two, publishing without approval; tier three, METR's embedded setup. Accenture and Faculty are also partners, with early talks including Hugging Face.

【Verifiable】

Anthropic announcement: https://www.anthropic.com/news/an-update-on-independent-evaluators

@AnthropicAI post: https://x.com/AnthropicAI/status/2110784662071398807

【Our read】

1. This puts METR's embedded-assessment proposal from a day earlier (e/879) into practice and answers the Reuters account of labs seeking control over evaluators' publications (e/862): the key terms are no veto and challengeable redactions.

2. The real test is METR's first report: how specific it is and how much is redacted will say more than the pledge.

3. Six people with carve-outs is still small, and withholding architecture by default limits independent judgment of training-related risk.

4. Coming after the FTC subpoenas (e/886) and EU AI Office requests, regulatory pressure is making outside evaluation an industry baseline. Watch whether OpenAI follows.