Ex-Anthropic hire and ex-METR COO raise $40M for AI agent audit startup AIUC

A startup founded by an early Anthropic employee and the former chief operating officer of AI safety evaluator METR has raised $40 million to build a third-party audit and certification layer for AI agents, TechCrunch reports. The company, Artificial Intelligence Underwriting Company (AIUC), announced the Series A on Tuesday, led by Ribbit Capital with First Harmonic participating. It previously raised a $15 million seed round backed by Nat Friedman through his fund NFDG, along with Emergence, Terrain and Anthropic co-founder Ben Mann, bringing total funding to $55 million.

Who is behind it

The founders are Rune Kvist, one of Anthropic's earliest hires, and Rajiv Dattani, who ran METR as COO from 2024 to 2025 and remains on its board. The two are also brothers-in-law. Their core argument, as Kvist put it to TechCrunch, is that intelligence and controllability pull in opposite directions: the smarter AI becomes, the harder it gets to deploy and govern.

Kvist said large institutions such as banks, hospitals, governments and militaries are no longer holding back on AI because models lack capability. They hesitate because they have made promises to their own customers about what a system will and will not do, and nobody in the market can currently guarantee that those promises hold.

A SOC 2 for AI agents

AIUC's answer borrows a familiar playbook from cybersecurity. Just as SOC 2 became the de facto trust benchmark for cloud vendors, AIUC has developed a standard called AIUC-1 along with a testing service that validates agents against it.

To shape the standard, the company assembled a consortium of roughly 250 security and risk leaders, the people who actually purchase agents on behalf of enterprises. Dattani told TechCrunch that AIUC meets with this group monthly and asks what they would want to see from an agent vendor before signing off on a purchase.

What the audit actually does

An agent under review is run through a suite of about 5,000 tests examining how it behaves in scenarios involving jailbreaks, hallucinations and data leaks. The output is a roughly 100-page report laying out where the agent performs safely and reliably, and where it does not.

Notably, AIUC uses AI agents to administer the tests and AI to analyze the resulting data, though humans verify the final audit before it ships, according to Kvist. The startup lists Cursor, Lovable, Harvey and ElevenLabs among its customers.

A crowded and urgent moment

The model has obvious echoes of Dattani's old employer. METR performs similar evaluations for frontier labs, though its public work has until recently concentrated on performance, meaning whether agents can reliably complete tasks rather than how they fail. METR was also one of the independent organizations OpenAI enlisted to examine its Hugging Face incident.

The backdrop is growing industry anxiety. Anthropic CEO Dario Amodei recently urged the field to slow the pace of frontier development, citing a rise in reports of misbehaving systems, and suggested frontier labs be required to host embedded third-party evaluators to verify safety claims, naming METR as a candidate. AIUC is not proposing to station staff inside customer organizations, but the underlying idea is the same: give buyers an independent read on whether an agent can be trusted before they commit. As Dattani framed it, the report tells a customer where the agent passes and is dependable, and where concerns exist that they should weigh before buying.

Why it matters

AI agents are moving from demos into workflows with real data, real money and real permissions, and enterprises currently have no widely accepted way to verify how one will behave under pressure. A certification layer modeled on SOC 2 could turn agent safety from a vague assurance into a procurement requirement, which is often how security practices become universal.

It also marks a commercial turn for AI safety talent: veterans of Anthropic and METR are now selling trust as a product rather than publishing it as research. The open questions are whether a 5,000-test battery can keep up with rapidly improving models, and whether certification matures into genuine assurance or settles into checkbox compliance. Either way, AIUC's $55 million in backing suggests investors believe agent audits are about to become a line item in every enterprise AI budget.