Anthropic says METR will publish its own analysis of Anthropic's risks, embeds six METR staff, and sets out three tiers of access for outside evaluators

Anthropic published “An update on independent evaluators” on 15 October. It says it has onboarded six METR employees with company hardware, office space and access to internal tools and systems that are roughly on par with what its own teams have.

The post describes three tiers. For the first, narrowly scoped investigations, it says evaluators “can publish their findings with our approval”. Second-tier evaluators “can publish without our approval”. At the third tier, “METR will publish its analysis of our risk levels, incidents, or practices.” Anthropic writes: “Anthropic can’t veto publication, but we can require them to redact information if it falls into narrow, predefined categories like the ones above.” Its examples include customer data, attorney-client privileged information and model architectures.

It names other evaluators. SecureBio is reviewing the chemical and biological chapter of Anthropic's August Risk Report. Accenture and Faculty are in what it calls an ambitious partnership, and talks with Hugging Face are in early discussions. “We’re embedding evaluators voluntarily, and we believe every frontier AI company should do the same,” the post says.

The post does not mention Reuters. On 14 October Reuters reported that Anthropic and OpenAI had sought control over what outside safety evaluators may publish; the link between the two is the Switchboard's, not Anthropic's. The arrangement is Anthropic's own description, and no evaluator has yet published under it.

• Confirmed Anthropic says it has onboarded six METR employees with company hardware, office space and access to internal tools and systems. Anthropic
• Claimed “METR will publish its analysis of our risk levels, incidents, or practices.” Anthropic says it can require redactions only in narrow, predefined categories. Anthropic
• Confirmed First-tier evaluators can publish with Anthropic's approval; second-tier evaluators can publish without it. Anthropic