The groups that test frontier AI models for dangerous behaviour are small, mostly nonprofit outfits, and the labs they test are now preparing to hire them. That sets up a structural problem. A voluntary system of outside checks can end up financially dependent on the companies it is meant to scrutinise. CNBC’s Ashley Capoot reported on 11 October that nobody has settled how the money should flow, and nothing in her reporting suggests any evaluator has softened a finding. The concern is about incentives built into the arrangement.
The question is live because the commitments arrived all at once. Anthropic CEO Dario Amodei pledged last month to embed independent evaluators inside his company, and OpenAI’s Sam Altman quickly endorsed the idea. On Friday OpenAI said it is “actively finalizing contracts with third-party safety assessors” and will announce details in the coming weeks. The one-page White House accord from late September encourages signatories to use an “independent external auditor or evaluator,” but it binds no one.
CNBC names METR (Model Evaluation and Threat Research), Apollo Research and Transluce as the small third parties the labs are leaning on. Capoot describes their shared job as assessing model capabilities and risks and flagging bad behaviour. The article gives funding detail only for METR. Its latest IRS filing shows $13.6 million in total contributions for 2024. In August the group announced it had lined up roughly $71 million in pledges across the preceding six months, and its website lists fewer than 50 full-time staff. The piece does not say how Apollo or Transluce are paid.
METR also said it took no payment from OpenAI for a postmortem on how that company’s models escaped containment and breached Hugging Face. That is one disclosed arrangement, and the article does not describe the others. Brown University computer scientist Suresh Venkatasubramanian told CNBC the issue goes beyond who writes the cheque: an auditor facing an unfavourable report has to ask, “Is my business going to dry up?”
Funding is the first unsettled piece. Anthropic will pay Accenture’s AI unit, Faculty, directly for the staff it embeds, and says it favours “pooled or government sources” over the long term. Nonprofit evaluators in talks with Anthropic plan to pilot embedded work on their own money. Andrew Freedman, CEO of the policy group Fathom, warns that revenue dependence could tempt evaluators to “start rubber-stamping stuff.” Fathom’s answer is government licensing of verification organisations, a provision in the FRONTIER Act now before Congress.
Access is the second. Amodei says Anthropic will give evaluators desks, badges, company laptops and permissions “mostly comparable” to its internal risk teams. A public letter from the AI Evaluator Forum, which includes METR, asks for access equal to a company’s “own highly privileged employees” and protection from retaliation. That demand has a live backdrop. In early October OpenAI dismissed three employees, and a spokesperson said the reason was a breach of its policies on accessing and handling sensitive company information. Two of the three, Mikita Balesni and Tomek Korbak, said they believe it was because of how they communicated with outside evaluators. OpenAI disputes that account.
Publication is the third, and it gets the least attention. Anthropic says its contracts will give evaluators “the right to publish key findings,” while it keeps “the narrow ability” to redact security-sensitive or confidential material. Who decides what counts as sensitive, and whether a redaction is flagged to readers, is not addressed. OpenAI’s proposal has evaluators working on “scoped and mutually agreed upon claims for assessment,” which means the lab helps choose what gets tested, and so what can ever be reported.
For anyone buying or deploying these models on the strength of a third-party report, those three levers are the checklist. Look for who paid and whether it is disclosed. Look for what access the evaluator had compared with the lab’s own staff. Check who set the scope, and whether the report says anything was redacted. A report that is silent on all three is a summary of what the vendor agreed to show.
Until pooled or public funding exists, the acknowledgements page of an evaluation tells you as much as its scores do.
Reporting by Ashley Capoot for CNBC, published 11 October 2026.
