NextAI+ Praxis--:----UTCIn this episode of Latent Space, released on 16 September 2026, Rune Kvist, Anthropic’s first product hire and co-founder of AIUC, advances a central proposition: the binding constraint on AI procurement is trust rather than capability. Businesses want to adopt AI, but hesitate because they do not know who will be held responsible when something goes wrong.
AIUC has just completed a $40 million Series A round co-led by Ribbit Capital and First Harmonic. Its customers include Cursor, Harvey, Lovable and ElevenLabs. At the start of the episode, Kvist makes the point directly: with the arrival of models such as Fable and Mythos, capability is no longer the issue; risk has become the binding constraint on procurement. This is now an immediate reality rather than a hypothetical concern.
AIUC-1 is AIUC’s standard for AI agent security, safety and reliability. Its design brings the concerns that keep enterprise security leaders awake at night into a comprehensive framework. The standard spans six categories, addressing issues such as jailbreak resistance, hallucination rates and data leakage, and requires quarterly updates. Updating every quarter rather than every decade is a core design principle, reflecting how much faster AI risks evolve than traditional standards are revised.
The standard has three types of requirements: technical controls (what must be in place), testing controls (what an independent third party must test), and policy controls (whose name is on the contract when something goes wrong). Certification takes three to ten weeks and is valid for one year, with quarterly retesting required to maintain it. The AIUC-1 consortium comprises security leaders from Fortune 1000 companies and meets twice each quarter to update the standard.
AIUC works with Lloyd’s of London, which has a 400-year history and has never declined a claim, to provide AI agent insurance to certified companies. ElevenLabs has purchased the first AI agent insurance policy. The logic is straightforward: enterprises gain the confidence to deploy when the world’s most conservative insurers are willing to put their balance sheets behind a system, rather than simply relying on a startup’s security blog.
Kvist identifies copyright as the AI risk that is currently hardest to underwrite. The reason is adverse selection: companies most interested in buying copyright infringement cover are often those that know most clearly that their training data contains copyrighted material. Insurers respond to that signal by tightening underwriting.
Kvist uses a specific hypothetical to frame the liability question: a user pays $20 for a Cursor subscription, uses it to write code, and that code causes an accident resulting in $200 million in losses. Who pays? He cites the Air Canada chatbot case as a precedent: the court held that putting AI in front of customers means allowing it to make legally binding commitments on the company’s behalf; the company cannot avoid liability by arguing that “the chatbot said it.” The existence of a standard directly affects how a court may subsequently assess whether a duty of care has been met.
AIUC’s expansion starts with agents, where AIUC-1 is already operational, then moves to the model layer as a trusted intermediary between governments and labs, and eventually to robotics, where physical AI creates greater liability when things go wrong. Kvist believes AIUC will remain relevant even if artificial general intelligence (AGI) is achieved, because there is one role labs can never perform for themselves: serving as their own watchdog.
The episode’s value lies in the clarity Kvist brings to a question that is often left ambiguous, rather than in the size of AIUC’s funding round: putting AI governance into practice ultimately depends on insurance pricing and legal precedents, rather than ethical declarations. Four dimensions of this argument merit closer examination.
Kvist repeatedly uses Waymo as an analogy. Autonomous driving technology has already been shown to outperform human drivers, yet four years later, passengers still cannot take a Waymo to the airport. He attributes the constraint to the absence of an institution able to explain to the public, governments and insurers who will be responsible if something goes wrong.
The same logic applies to enterprise AI deployment. Demos are easy to sell: capabilities keep improving and demonstrations become more impressive. But when a hospital or bank considers full deployment, the central challenge becomes accountability for risk. The point at which procurement stalls is whether the person approving deployment can withstand the consequences of a failure, rather than simply whether the AI works.
AIUC-1 is updated quarterly. That cadence reflects an explicit position: AI risks evolve at a pace unlike that of earlier technologies. As Kvist puts it, the leading security concerns three months ago are no longer the same as those of today.
This has direct policy implications for AI governance. Traditional standards systems, including ISO and NIST, operate on revision cycles measured in years or even decades, implicitly assuming that technological risks accumulate gradually. Generative AI’s risk landscape, by contrast, is being rewritten every quarter. A standard that cannot keep pace with evolving risks offers enterprises a false sense of security rather than meaningful protection.
Kvist’s analysis of copyright is among the episode’s most insightful. The underwriting difficulty arises from a textbook case of adverse selection: companies with greater copyright exposure have stronger incentives to buy insurance, and that demand makes insurers more cautious. Pricing alone cannot automatically resolve the cycle.
The deeper problem is information asymmetry. Downstream users lack visibility into the training data of open-source models. They cannot determine whether they are using a model “contaminated” by copyrighted material, and insurers lack the information needed to price that risk. The difficulty of resolving copyright risk exposes a broader lack of provenance across the AI supply chain: when the chain of responsibility cannot be traced, risk cannot be allocated appropriately.
Towards the end of the episode, Kvist offers a proposition that warrants consideration in its own right: labs can never serve as their own watchdogs. His reasoning concerns incentives rather than a lack of concern. Competitive pressure can encourage labs to withhold information; in his account, this follows from the market structure rather than individual moral failings.
He invokes a familiar scene from The Big Short: if Moody’s will not award an AAA rating, the issuer can turn to S&P, creating a “race to the bottom” among competing oversight providers. His answer is to bring insurers into the process. Insurers are the only participants that must actually pay after a loss, making them the only participants with an incentive to resist that race. An oversight provider that must pay the claims itself has little reason to understate the risk.
The episode raises a question about governance infrastructure in the AI era that goes beyond insurance products and funding milestones: before AI systems become deeply embedded in enterprise operations, who defines “safe enough,” who verifies that assessment, and who pays when something goes wrong?
Lab self-regulation cannot resolve this problem, and government legislation cannot keep pace with it. Kvist’s answer is to fill the gap through market mechanisms: standards define the boundaries of risk, independent testing verifies actual conditions, and insurers put their balance sheets behind the assessment.
What AI governance ultimately needs is not more statements of principle, but infrastructure that assigns responsibility for harmful outcomes and verifies positive commitments. That infrastructure is being assembled, piece by piece, from insurance, security auditing and legal precedents.
Cite as · Deep Reads · 22 October 2026
If you want both columns delivered together, four times a year, in one quiet email — leave an address. Otherwise just bookmark this page.