New startup offers AI safety audits to help companies tame risky agents

As AI systems grow more powerful, organizations are increasingly hesitant to deploy them—not because the technology lacks capability, but because they cannot guarantee what an agent will actually do once it's in production. Two founders with deep roots in AI safety think they have a solution.

Rune Kvist, an early employee at Anthropic, and Rajiv Dattani, former chief operating officer of the AI safety research organization METR, have launched Artificial Intelligence Underwriting Company (AIUC) to provide independent audits and certification for enterprise AI agents. The pair say their work addresses a critical gap: enterprises deploying AI need assurance that agents won't misbehave, leak data, or circumvent safeguards in ways they've promised their own customers won't happen.

The funding and the business model#

AIUC announced a $40 million Series A round led by Ribbit Capital and First Harmonic on Tuesday. Combined with a $15 million seed round closed earlier from investor Nat Friedman, Anthropic co-founder Ben Mann, and others, the startup has raised $55 million to date. Early customers include code editor Cursor, design tool Lovable, legal AI platform Harvey, and voice company ElevenLabs.

The company's approach draws inspiration from established cybersecurity practices. Rather than inventing something entirely new, AIUC adapted the widely trusted SOC 2 framework used in security compliance—creating an analogous standard called AIUC-1 specifically designed for evaluating AI agents.

How the testing works#

To shape their standard and testing protocol, the founders gathered input from approximately 250 security and risk leaders at organizations considering AI adoption. These individuals become the de facto buyers and decision-makers, and their feedback directly informed what the certification process would measure.

The startup puts each agent through a battery of roughly 5,000 tests, simulating scenarios that probe for common failure modes:

• Jailbreaks that try to override safety instructions
• Hallucinations and false claims
• Unintended data leaks or privacy breaches

The testing generates a detailed 100-page report mapping where an agent performs reliably and where vulnerabilities exist. Notably, AIUC relies on AI agents themselves to run many of the tests and process the resulting data—though human auditors verify the final certification, Kvist emphasized.

Why this matters now