AIUC raises $40M to put an insurance policy behind its AIUC-1 agent audit

An audit and a policy are two different products, and they fail differently. The audit says an agent was tested against a written control set. The policy says someone pays when the agent causes a loss anyway. Of Kvist, who was Anthropic's first product hire, latent.space writes that he "is betting that the biggest constraint on AI adoption won't be capability it will be trust" ^[3][15]. For the policy half to price anything, the wording has to define a covered loss, and the underwriter needs some view of frequency and severity for a failure class with almost no claims history.

Severity is where the episode starts. One of its listed segments asks what happens if a $20 Cursor subscription contributes to a $200M plane crash ^[6]. A liability cap pegged to fees paid does not reach that loss. Divide one into the other and it is ten million times the monthly contract value ^[1].

The episode also argues that AI standards may need to update every quarter instead of every decade ^[7]. Four revisions a year against one every ten years is forty times as many revision events ^[2]. Enterprise compliance runs on annual audit cycles, so a control set that moves quarterly means the vendor re-tests four times a year and the certificate a buyer reads is at most one revision from stale.

There is a premise in the same episode that limits what any certificate can say. Kvist argues that every model can ultimately be jailbroken ^[8]. Under that premise an audit cannot certify the absence of a jailbreak; it can only report measured resistance on the adversarial cases someone wrote. The episode's own complaint about the field is that most AI companies optimize the happy path without seriously stress-testing adversarial cases ^[9].

The announcement itself lives in a podcast description. Swyx opened by saying "we're in the studio with Rune from AIUC, the Artificial Intelligence Underwriting Company", and the transcript published on the page stops mid-word a sentence later ^[10][17]. That transcript is where the public record of this round currently ends. AIUC did not disclose premiums, coverage limits or the AIUC-1 control list, and the description says the company works with Cursor, Harvey, Lovable and ElevenLabs without saying what any of them bought ^[20]. Lloyd's of London comes up as a discussion topic on how AI systems can be insured ^[12]. The company's earlier backing came from NFDG ^[11].

One item on the episode's agenda undercuts what the audit measures: models becoming aware they are being tested ^[14]. An eval-aware model makes the audit a measurement of behavior under observation, while the underwriter is pricing behavior in production.