AIUC raises $40M to aim its 5,000-combination agent audit at frontier models

The agent cleared its pilot and then stopped moving. Rune Kvist, AIUC's co-founder and chief executive, said most enterprises have a list of agents that were approved in pilots but "stalled at the security review" ^[4]. Proof of security and reliability, he added, has become the main bottleneck ^[5]. What the company sells into that stall is a document a risk owner will accept.

The document has a shelf life. One certification runs the agent through about 5,000 combinations of risk and attack chosen for the type of business deploying it ^[6], with jailbreaks, hallucinations and data leaks among the failure modes tested ^[7]. AIUC audits each agent independently and recertifies every quarter as attack techniques change ^[8]. Four audits a year at that size works out to roughly 20,000 combinations annually for a single agent ^[20]. It also means the certificate in the file describes the threat picture as of the last audit, up to about three months back ^[21]. Much of the testing is done by AI agents, with humans verifying the final audit, according to TechCrunch ^[9].

Rajiv Dattani, the co-founder who was a partner in McKinsey's insurance practice and then chief operating officer of the model evaluation nonprofit METR ^[14], described the sequence in older terms. "When electricity was burning down houses, the insurers paying the bill funded Underwriters Laboratories to test and certify products," he said. "To this day, the UL mark is on most light bulbs across America. AI needs the same combination of standards, testing and insurance." ^[15]

A consortium of more than 250 security and risk leaders from Fortune 1000 companies shapes the standard and pushes for its adoption inside their own organizations ^[12]. Certificate holders include Anysphere, the developer of Cursor, plus Harvey AI, KPMG, Lovable Labs, UiPath and Fin, the customer service software maker formerly called Intercom ^[11].

AIUC's own argument is that uncertainty over risk now slows adoption more than any shortfall in what the technology can do ^[3]. The company also argues that the security reviews which blocked application rollouts first are now building around the models themselves ^[16]. Ribbit Capital led the Series A with participation from First Harmonic ^[17]. That follows the $15 million seed that Nat Friedman at NFDG led when the company launched in July 2025, for $55 million raised in total ^[18]. The two rounds are about 14 months apart ^[22]. The report does not state what certification costs ^[23]. Micky Malka of Ribbit said AI is heading down the same road as financial services, "and it is moving faster than the systems companies use to evaluate it." ^[19]

For the person who has to move a blocked pilot, the queue sorts on who the reviewer is and what the agent can do. Does a named risk owner in your company treat third-party attestation as sufficient, or do they want your own test results? And can the agent take an action nobody can reverse, such as a payment or a statement to a customer that becomes the company's position? Where attestation is accepted and the actions are reversible, a certificate is the cheapest way to get the signature. Where the actions are not reversible, the audit is a prerequisite for coverage, and the coverage is what Eleven Labs said it went and bought in February ^[10]. Where the reviewer wants their own evidence, buying the standard adds four renewal dates a year to your calendar and leaves the internal testing where it was ^[8].