AIUC raises $40M to build a safety certification layer for AI agents

As AI agents move from chat interfaces into business workflows, the central enterprise question is changing. Companies are no longer asking only whether a model is capable enough to complete a task. They also need to know whether an agent will respect boundaries, protect sensitive information and behave predictably when instructions go wrong.

An audit model for agent risk

Artificial Intelligence Underwriting Company, or AIUC, was founded by Rune Kvist, an early Anthropic employee, and Rajiv Dattani, the former COO of AI safety organization METR. The startup has announced a $40 million Series A led by Ribbit Capital, with participation from First Harmonic. It had previously raised a $15 million seed round, bringing total funding to $55 million. AIUC identifies Cursor, Lovable, Harvey and ElevenLabs as customers.

Kvist’s argument is that increasing model intelligence does not automatically make adoption easier. More capable systems can also be harder to control. For banks, hospitals, governments and other organizations, the obstacle may be less about raw performance than about whether they can make credible promises about what an agent will and will not do.

How the testing works

• AIUC-1 framework: AIUC has created a standard inspired by SOC 2, a widely used cybersecurity compliance model, but aimed at AI agents.
• Buyer-led requirements: The company says it has assembled a consortium of about 250 security and risk leaders. Their concerns help shape the questions and tests used in the framework.
• Large test suite: An agent is put through roughly 5,000 tests involving scenarios such as jailbreak attempts, hallucinations and data leaks.
• Detailed reporting: The process produces a report of about 100 pages describing areas where the system appears reliable and areas that require caution. AI agents help run the tests and analyze results, while humans verify the final audit.

The approach overlaps with some of METR’s work, although METR has largely focused on measuring whether frontier models and agents can reliably complete tasks. AIUC is positioning its service closer to the enterprise purchasing decision: an outside assessment that helps a customer understand where an agent can be trusted and where safeguards may be needed.

Why it matters

A conventional benchmark can show that an agent completes a task under controlled conditions. It may not reveal whether the same system will expose data, bypass a restriction or make an unsafe decision when connected to real tools and business information. A standardized audit could turn those concerns into comparable procurement criteria and give developers a structured way to identify weaknesses before deployment.

There are important limits. No fixed test suite can cover every combination of permissions, data and instructions. An agent’s risk profile can change when its tools or operating environment change, and certification should not be mistaken for a guarantee of absolute safety. AIUC will also need to demonstrate that its tests are transparent, repeatable and sufficiently independent from the companies it evaluates.

The broader shift is significant: AI safety is moving from research-lab evaluation toward operational governance. AIUC’s funding reflects demand for that layer, but whether AIUC-1 becomes something like SOC 2 for agents will depend on adoption by buyers, developers, regulators and independent researchers.

Source: TechCrunch AI