Keeping AI agents in check: Kyvvu raises €1 million

Dutch startup Kyvvu has raised €1 million to fix a security gap that most companies deploying AI agents don't yet realize they have: an agent can follow every rule it's given and still cause a breach, simply by combining permitted actions in the wrong order.

The pre-seed round was backed by Volta Ventures and the Brabant Development Agency (BOM). Kyvvu, based in Den Bosch, builds a security layer that sits inside AI agents and, in under a millisecond, checks whether each action they're about to take should actually go through.

The problem Kyvvu is targeting has become more urgent as companies move from using AI to suggest actions to letting it perform them, such as updating customer records, moving data between systems, issuing refunds, or running entire back-office processes without a human in the loop. A wrong suggestion is a minor inconvenience. An incorrect action taken autonomously is an incident, and in sectors like banking, insurance, and healthcare, it's often one that must be reported to a regulator.

Taming AI agents

According to Kyvvu, the real danger isn't in what an agent is technically permitted to do. Rather, the risk lies in the order it chooses to do things in. An agent might be allowed to read a certain customer record, and separately allowed to send emails outside the organization. Individually, both actions are approved. Combined, in the wrong sequence, they amount to a data leak. Because agents decide their own path while running, that harmful combination can slip past security tools that only check one action at a time.

"The internet only scaled once we built the security layer that made it safe to run real workloads on it. Agents are at exactly that point," said Jeroen Ghijsen, co-founder and CEO of Kyvvu. "They will only be trusted with work that matters once something inside them is checking each action before it happens."

Kyvvu's product, called the Agent Security Kernel (ASK), sits inside the agent itself and evaluates every action — reading data, calling a tool, sending a message — against everything the agent has already done in that task, before deciding to allow, warn, or block it. The rules are fixed, version-controlled policies that a compliance officer can inspect, rather than judgments made by a second AI model, and they never leave the customer's own systems. The company says the engine is deterministic, meaning the same input always produces the same verdict, and decisions are returned fast enough that the agent's performance isn't slowed down.

The need for a control point

The kernel is framework-agnostic and available as source code, and it's already running at financial services, insurance, and healthcare organizations in the Netherlands. Because every action passes through the same checkpoint, the system also generates a detailed record of what each agent did — something Kyvvu says lines up with record-keeping and human-oversight requirements under the EU AI Act and GDPR.

"Enterprises are deploying agents faster than they can govern them, and the tooling to keep those agents in line simply has not existed," said Sander Vonk, Managing Partner at Volta Ventures. "We think a runtime control point becomes a hard requirement for running agents at scale, the way endpoint security did before it. We are backing Jeroen, Maurits and the team to define what that looks like."

Keeping North Brabant competitive in AI

The company grew out of research at the Jheronimus Academy of Data Science in Den Bosch, where the founding team was building AI agents before most enterprises had deployed one in production. Kyvvu will use the new funding to expand its engineering team, grow its partner network, and support deployments across Europe.

"Kyvvu is the kind of company that keeps this region competitive in AI," said Robin Hendrickx, Senior Investment Manager and Team Lead Key Technologies at BOM. "Every large enterprise putting agents into production is about to run into the same wall, and this team is already on the other side of it, from Den Bosch, with live customers in some of the most heavily regulated sectors in the country."