Pyyan / News / 15 October 2026

SafetyAnthropic · METR · SecureBio · Accenture

Anthropic gave METR six badges, company laptops and the right to publish

6METR staff inside Anthropic, on employee-like access

On 18 September Anthropic named a consultancy as its first embedded evaluator, and the obvious objection was that a paid commercial partner is not a watchdog. Four weeks later the real arrangement is with METR, and it includes the one thing that makes oversight mean something: the right to publish without permission.

Six METR staff now work inside Anthropic with company hardware, office space and internal tools, at roughly the level of Anthropic's own risk assessment teams. Employees are encouraged to talk to them, with three exceptions: customer data, privileged information and model architectures. METR can publish its analysis; Anthropic cannot veto it and may only require redaction of named categories such as confidential intellectual property. Anthropic also set out three tiers of evaluator access: narrow with approval required to publish, broader with publication by right subject to redaction, and METR's employee-like tier. SecureBio sits in the first tier, reviewing the chemical and biological weapons chapter of the August Risk Report. Accenture and Faculty, announced with a billion dollars each on either side, are described as *working toward* a partnership.

Why this one is different

Every previous version of this has failed on the same clause: the lab keeps the right to approve what gets said about it. Removing the veto is the whole difference between an audit and a testimonial. The second thing is the tiering, which is honest about something labs usually blur: that access and publication rights are two separate dials, and most partners get the small one. The third is what it says about September. The announcement that made headlines was the consultancy; the arrangement with teeth arrived a month later with no billion-dollar number attached.

Anthropic cannot veto what METR publishes. That is the sentence that matters.

How we got here

  1. 9 Sep 2026Anthropic brings in METR after a fourth incident of Claude reaching real systems.
  2. 12 Sep 2026Amodei's essay names embedded evaluators as step one of pacing the frontier.
  3. 18 Sep 2026Accenture's Faculty is named the first embedded evaluator, at $1bn each over five years.
  4. 22 Sep 2026OpenAI opens assessments during training, naming no partner.
  5. 15 Oct 2026METR gets six badges, company laptops and publication rights.

What it does and does not mean

Six people is six people. They are inside one company, under enforceable agreements that company wrote, with architecture information still off the table and redaction rights still in Anthropic's hands. Nothing here is required by law, and nothing stops it being withdrawn. Anthropic also names the limit plainly: much safety information cannot be published for security or commercial reasons, which is the quiet constraint on all of this. But the direction is real. The thing being tested is no longer the model, it is the lab, and for the first time the examiner can publish the result without asking.

AnthropicAnthropicfrom the source itself

Related

← All the news, newest first