Benchmarks · Benchmark
Agents Get a Trustmark
AIUC-1 is trying to be SOC 2 for agents: thousands of evals, a certificate, and now a $40M Series A to sell audits and insurance. A badge is not a kill switch.
Alex Rivera · Sep 17, 2026 · Updated Oct 23, 2026
On this page
Frontier labs spent last weekend arguing about embedding evaluators inside the companies that train models. AIUC is selling the other product: a certificate for the agents those models become once they hit production.
On 15 September 2026, co-founders Rajiv Dattani and Rune Kvist announced a $40 million Series A led by Ribbit Capital, with First Harmonic and Terrain. Seed from NFDG brings the total to $55 million. The company name is the pitch: Artificial Intelligence Underwriting Company. Their analogy is Underwriters Laboratories: electricity burned houses until insurers funded a standard and a mark.
Image: AIUC, Series A announcement, 15 September 2026.
What AIUC-1 actually tests
AIUC-1 is their agent standard. It covers data and privacy, security, safety, reliability, accountability, and societal risk. The technical story they sell: on the order of 5,000 risk-and-attack combinations, tuned per business type: jailbreaks, hallucinations, data leaks. Agents help run the tests; humans sign the audit. Technical tests re-run at least quarterly. The full certificate is annual, like a SOC 2 display term.
Unlike SOC 2 Type II, they claim forward-looking controls: log review, adversarial tests, not only a look back at last year's samples. Fail to renew and the mark is supposed to come down. Material holes mean a qualified or adverse report until P0/P1 issues are retested.
Certification typically takes four to eight weeks with an accredited auditor. The certificate attaches to specific products, not the whole company is safe. That distinction matters the first time a vendor puts the logo on a homepage while a different agent in the same org is untested.
Who already wears the mark
AIUC names Cursor, ElevenLabs, Harvey, KPMG, Lovable, UiPath, and Fin as certified or in the program. Cursor's coding agents reportedly went through thousands of evals across 12 categories, including coding-agent-specific requirements published in Q3 2026. Harvey's legal platform took 3,063 evals; AIUC says no critical or major vulnerabilities. KPMG is the first Big Four firm they cite, for an aIQ Capture agentic platform. ElevenLabs is using the cert as the hook for agent insurance.
TechCrunch's 15 September piece frames the same founding pair as an answer to rogue enterprise agents, naming Cursor, Lovable, Harvey, and ElevenLabs as customers. The Series A post is more insurance-and-standards than we will stop a swarm. Hold both in your head.
Where this sits next to lab watchdogs
Amodei's embedded evaluators want desks inside Anthropic. AIUC wants a report you can attach to a purchase order. One looks at training pipelines. The other looks at deployed agents. Neither can halt a training run. AIUC is explicit that it is not embedding at customer sites. Publish-and-certify is still the enforcement.
For a founder buying an agent that writes production code or talks to customers, the operational questions are: which product version is on the cert, when the last quarterly eval ran, and whether a failed retest is public. A trustmark without those dates is SOC 2 theater with a new acronym.
Sources
AIUC, Series A to build the confidence infrastructure for frontier AI, 15 September 2026. https://aiuc.com/updates/series-a-announcement
AIUC-1, The standard for AI agent security, safety and reliability. https://www.aiuc-1.com/
TechCrunch, Early Anthropic hire, former METR COO have found a way to rein in rogue AI agents, 15 September 2026. https://techcrunch.com/2026/09/15/early-anthropic-hire-former-metr-coo-have-found-a-way-to-rein-in-rogue-ai-agents/