Kyvvu

Kyvvu is developing an Agent Security Kernel (ASK): an in-process security layer that evaluates every action of an autonomous AI agent against the full path of the executed task, and deterministically allows, warns or blocks that action — in less than a millisecond, without data leaving its own infrastructure. The approach is based on the reference monitor concept used in operating systems to confine processes, now applied to AI agents.

The core of the distinction is path-dependent enforcement: an individual step can be legitimate, but the same step after reading personal data and calling an external system can constitute a data leak. Kyvvu detects such dangerous sequences and blocks them before they are executed. Policies are managed as version-controlled YAML manifests in the customer's own git environment. The platform integrates with Python, LangChain, LangGraph, Microsoft Copilot Studio and other frameworks.

Kyvvu focuses on regulated sectors in the Netherlands, including financial services, insurance and healthcare. As a European company, the approach — in which no data leaves its own environment — naturally aligns with requirements around data sovereignty and the EU AI Act (Art. 9, 12, 14 and 15). The approach is scientifically substantiated in the arXiv paper "Runtime Governance for AI Agents: Policies on Paths".