Anthropic's first embedded safety evaluator is Accenture -- the same company it just trained 30,000 staff on Claude

Anthropic and Accenture announced Thursday that they will each invest at least $1 billion over the next five years to embed a team of Accenture evaluators inside Anthropic, with access the companies describe as "comparable to an employee's." Drawn from Faculty, the AI consultancy Accenture acquired, the team will watch models take shape during training, follow the decisions that govern how they're built and deployed, speak directly with staff, and report incidents -- work that, until now, third-party evaluators have done from outside, looking only at a finished model's outputs. Anthropic frames the shift as closing a blind spot traditional red-teaming can't reach: assessing how the company operates and whether it's keeping its own safety commitments, not just what a released model does when prompted.

The deal is a direct execution of a plan Anthropic CEO Dario Amodei published six days earlier. In "We Must Pace the Frontier," Amodei proposed a three-step framework for slowing frontier AI development without halting it -- starting with a unilateral commitment to grant outside evaluators "desks in our offices, access badges, and company laptops," with permissions "mostly comparable to what internal risk assessment teams have." The Accenture terms match that description almost line for line, down to the promise that evaluators can publish risk findings "without editorial control by Anthropic," subject only to narrow redactions for security, legal or confidentiality reasons. Sam Altman and Elon Musk both publicly endorsed Amodei's call within days of its publication.

Anthropic is the first frontier lab to turn Amodei's proposal into an actual partner. Sam Altman reposted the essay and pledged OpenAI to "the same practice" within days of its publication, but as of this announcement OpenAI has named no evaluator, set no timeline, and hasn't said which systems or data an eventual evaluator could access -- when reporters asked both companies to fill in those details in mid-September, neither answered. That gap is the plainest evidence the Accenture deal is a real commitment and not just a shared talking point: one lab has a signed partner and a dollar figure, the other has a repost.

Anthropic said the partnership is non-exclusive and that it is "in dialogue with METR and other nonprofit evaluators to pilot elements of embedded evaluation using their own funding" -- organizations with no commercial stake in Anthropic's business. That distinction went to the center of an open letter published the same day the Accenture deal was announced. Organized by the AI Evaluator Forum and signed by more than 130 researchers and practitioners, including Geoffrey Hinton, Stuart Russell, Arvind Narayanan, Joy Buolamwini and Yejin Choi, the letter set five minimum conditions for a credible embedded evaluator. The first: the evaluating organization "should be meaningfully independent," with full editorial control over its findings and no other significant commercial relationship with the company it evaluates.

Accenture does not clear that bar. In December 2025, the two companies launched the Accenture Anthropic Business Group, a multi-year partnership to train roughly 30,000 Accenture professionals on Claude and make Accenture what both companies called its largest Claude Code deployment. In March 2026, they went further, launching Cyber.AI, a joint cybersecurity platform that runs Claude as its reasoning engine -- which Accenture has since used to secure 1,600 of its own applications and more than 500,000 APIs. None of that history appears in either company's announcement of the evaluator partnership.

“AI should be safe by design, not safe by accident.” — Dr. Marc Warner, Accenture CTO and Faculty CEO

For now, the terms that would actually settle the independence question -- whether Accenture's evaluators get the editorial-control-free publishing rights Amodei's own essay promises, and whether their access is genuinely equivalent to a senior Anthropic employee's or something narrower -- remain unconfirmed by either company. (Anthropic said "many of the details about how it will operate are still being worked out," per its own announcement -- itself a reason to treat this as a framework agreement rather than a finished program.) The letter's other four conditions read like a checklist against exactly this kind of deal: multiple evaluators covering different risk areas rather than one exclusive vendor (Anthropic's non-exclusivity claim partly answers this); transparency about methods, findings and access terms, with only narrow redactions; protection from retaliation, including guaranteed funding that can't be cut after an unfavorable finding; and access equivalent to a company's most senior internal risk staff, not a limited guest account. Each is answerable in principle -- Anthropic could publish the access terms, the redaction policy, and the funding structure -- and none of it has been published yet.

The test that actually matters plays out over the next year, not in this week's terms. Anthropic says the arrangement makes its safety claims more verifiable rather than simply trusted on its word; the AI Evaluator Forum's letter says verifiability requires the independence guarantees this deal hasn't yet spelled out. Both can be true until Accenture's team publishes something. A finding Anthropic didn't want made public would answer the question this announcement leaves open; a year of silence would answer it too, just less usefully.