What exactly can METR see and publish?

METR staff get Anthropic devices, office space and access to internal tools and systems roughly equal to Anthropic's own risk evaluation team. They publish analyses of Anthropic's risk levels and incidents, and can also publish their view on whether any redaction was justified.

What is not shared with METR?

Anthropic kept a small number of categories out of the access, including customer data and information on model architecture.

How does METR's evaluation work?

METR proposed to frontier AI companies on October 14 that its staff access data on company devices, interview employees individually, and publish a preliminary report within weeks. Access is left to each company's discretion, so it will not equal that of senior internal staff.

AI Safety & AlignmentAI Business & IndustryITmedia AI＋3m agoOct 16, 2026, 13:00 JST

Anthropic gives METR staff employee-level access, no veto

Send on LINE

3 Key Points

What happened
Anthropic said on October 15 it has started giving METR's six staff company devices, office space and access to internal tools and systems, roughly matching its own risk team.
Why it matters
METR can publish its own analysis of Anthropic's risk levels and incidents, and Anthropic cannot block it — it can only ask for redactions in preset categories.

WHO IT HITSAI safety and risk teams at frontier model developers — and the outside evaluators they bring inside their walls — now have a concrete model for how much access an external reviewer gets, and how much veto the company keeps. Anthropic says it wants all frontier AI companies to do the same, so compliance and policy staff at rival labs may face pressure to match the arrangement.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

Anthropic frames this as the top step of a three-tier design for working with evaluators. METR sits at tier three, the highest. SecureBio is at tier one, where Anthropic's approval is required for publication, and it reviewed the chemical and biological weapons chapter of an August risk report. Anthropic is also working with Accenture and Faculty, and is in early-stage discussions with others including Hugging Face.

The arrangement follows CEO Dario Amodei's September essay "We Must Pace the Frontier," which laid out the resident-evaluator idea that the October 15 announcement puts into practice. METR opened its own proposal a day earlier, on October 14, aimed at all frontier AI companies, with the stated aim of giving the public more material to judge the risk that developers lose control of their internal AI. METR itself notes the proposal's limits: because access is left to each company's discretion, its staff will not be equivalent to senior internal employees.

OpenAI CEO Sam Altman endorsed Amodei's September essay and said his company would carry out a similar effort. On October 8, when OpenAI said it had fired three safety researchers, it explained that it was finalizing a contract with an external safety evaluation organization and would announce details within weeks.

FAQ

What exactly can METR see and publish?: METR staff get Anthropic devices, office space and access to internal tools and systems roughly equal to Anthropic's own risk evaluation team. They publish analyses of Anthropic's risk levels and incidents, and can also publish their view on whether any redaction was justified.
What is not shared with METR?: Anthropic kept a small number of categories out of the access, including customer data and information on model architecture.
How does METR's evaluation work?: METR proposed to frontier AI companies on October 14 that its staff access data on company devices, interview employees individually, and publish a preliminary report within weeks. Access is left to each company's discretion, so it will not equal that of senior internal staff.

ITmedia AI＋Read Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.