Third-party audit for AI agents: Anthropic and METR alumni raise $40 million

Banks, hospitals and defense agencies are not holding back on AI deployment because the models lack intelligence. They are holding back because no one can guarantee an agent will do only what it was promised. Rune Kvist, an early Anthropic employee, and Rajiv Dattani, former COO at the safety-research organization METR, founded Artificial Intelligence Underwriting Company (AIUC) to fill that gap: a third-party audit and certification layer for AI agents, modeled on the SOC 2 standard from cybersecurity.

On Tuesday the company announced a $40 million Series A led by Ribbit Capital with participation from First Harmonic. The round follows a $15 million seed from Nat Friedman via NFDG, alongside Emergence, Terrain and Ben Mann, a co-founder of Anthropic — $55 million in total. The customer roster already includes Cursor, Lovable, Harvey and ElevenLabs, signaling genuine demand from the companies building and selling agents to enterprises.

To build the AIUC-1 standard, the founders assembled a consortium of roughly 250 security and risk leaders — the buyer side. Monthly sessions with that group produced the requirement set: what to test, which questions to ask, which scenarios must be covered. The result is a suite of about 5,000 tests that run the agent against jailbreak attempts, hallucinations and data-leakage vectors. The final audit report runs to roughly 100 pages and maps where the agent is safe and where it is not.

Ironically, the thousands of tests are executed by AI agents themselves, and the results are analyzed by a language model. Humans step in only for final verification of the report, Kvist said. The approach enables scale and speed while keeping the binding sign-off in human hands — a practical compromise between broad coverage and legal liability.

Dattani served as COO at METR from 2024 to 2025 and remains a board member; the organization runs similar evaluations for frontier labs, though its focus until recently was on capabilities — whether an agent completes tasks — rather than safety alone. Anthropic chief executive Dario Amodei recently called for the industry to slow frontier development and embed third-party evaluators, naming METR as a possible candidate. AIUC does not offer physical embedding at the customer, but the principle is similar: give an organization an independent assessment before a purchase decision.

"Here is where it passes and you can trust it, and here is where there are concerns — be aware of this before you buy," Dattani summarized. AIUC does not promise an agent will never behave unexpectedly; it provides a transparent map of known risks. For enterprises stalled at the pilot stage by compliance and regulatory requirements, that may be the difference between "no" and "let's try in a controlled environment."