Blog
Notes on agent security.
Research, opinions and incident write-ups from the Kyvvu team.
Latest
Do AI agents need an identity?
Identity works for people because the person behind it changes slowly. An agent's model and tools keep changing, so what does an agent identity tell you about what it will do? Perhaps state the rules it runs under instead, and check them in the harness.
· Maurits Kaptein
Read it- blog
· Kyvvu
EU AI Act Compliance
How policies on paths give you the runtime evidence the AI Act requires
- blog
· Maurits Kaptein
NVIDIA's OpenShell and the harness: what each layer sees
NVIDIA’s OpenShell puts the agent process in a kernel-enforced sandbox. A sandbox governs the process; a monitor in the harness governs the action. How the two relate, and where SHADI and Sentry fit.
- blog#dora
· Maurits Kaptein
What does DORA ask of your AI agents?
DORA does not explicitly mention AI, yet every agent a bank or insurer builds, buys or hands to its developers is an ICT system to which DORA applies. In this post we try to analyze the DORA requirements and highlight how they relate to agent governance and security.
- blog#agent-security
· Maurits Kaptein
We Used to Be the Only Ones. We're Not Anymore.
A reflection on OWASP’s Agent Control Standard, Microsoft’s Agent Control Specification, and AGNTCY’s SHADI, seen through the lens of a reference monitor — and how SHADI extends what Kyvvu covers.
- news#kyvvu
· Maurits Kaptein
kyvvu-claude: policy enforcement for Claude Code, and a beta for coding agents
kyvvu-claude checks every Claude Code tool call against your policies before it runs, using Claude Code’s own hooks. Claude Code works today; Cursor, Copilot, Codex and OpenCode are being built with the beta teams.
- news#kyvvu
· Maurits Kaptein
Kyvvu raises a €1M pre-seed from Volta Ventures and BOM
We raised a €1 million pre-seed round from Volta Ventures and the Brabant Development Agency. What the round pays for: the team, the partner network, and security accreditation.
- blog#agents
· Maurits Kaptein
On the OpenAI–Hugging Face incident: what was doing the escaping?
A structural read of the OpenAI / Hugging Face agent incident: what actually broke, why ‘the AI escaped’ is not the most useful framing, and why one of the missing controls is a path-aware check on the harness.
- blog#ai-governance
· Maurits Kaptein
Building Secure Agents at Scale
We wrote a white paper on agent security: what AI agents are, what actually makes them dangerous, and the full menu of ways to secure them. Most of the danger turns out to be old and well understood. One thing is new, and that is where the difficulty lives.
- blog#ai-governance
· Maurits Kaptein
Where Does Your Agent's Data Go?
We spend a lot of our time arguing about what an agent is allowed to do. However, this question is quite intimately related to the data that the agent has access to, and where it is retrieving and sending that data. This post tries to sketch a part of the ‘canonical’ agent data architecture and its impact upon agent security.
- blog#ai-governance
· Maurits Kaptein
Agent Safety Is a Frontier, Not a Guarantee
“Safe”, “secure”, and “governed” sound like properties an agent either has or doesn’t. For agents they aren’t. Safety is a position on a frontier between usefulness and risk, no unconditional guarantee holds, and every governance method is a way of moving the frontier outward. Here we try to discuss the various approaches and explain where kyvvu sits.
- blog#ai-governance
· Maurits Kaptein
"Guardrails" Now Means Everything. Here's the Map It's Hiding.
The word “guardrails” has come to cover content filters, identity scoping, runtime monitoring, prompt hardening, and access control — all at once. When a term means everything, it specifies nothing. So we wrote down the actual map: every layer of the agent-security stack, what it really sees, where it stops, and why Kyvvu sits where it does.
- blog#ai-governance
· Maurits Kaptein
Everyone Shipped a Governance Layer This Month. They're All Stateless.
Microsoft and AWS both shipped agent-governance engines this spring, and the phrasing is nearly interchangeable. But read into the engines and they share one property we deliberately built against: they’re stateless. Here’s why the failures that actually hurt are individually innocent — and why that makes them a question about the path, not the action.
- blog#ai-governance
· Maurits Kaptein
Nine Seconds, Two CVEs, and a Three-Axis Vocabulary
Two recent incidents — an AI coding agent that deleted a production database in nine seconds, and two RCE vulnerabilities in Semantic Kernel — fail along the same axis. Both succeed if the runtime has no vocabulary to refuse in. Here is what we’ve landed on as the vocabulary it needs.
- blog#runtime-enforcement
· Maurits Kaptein
The Hot Path Tax: Why Runtime Governance Has to Be Sub-Millisecond
Most of the past week at Kyvvu was spent on a refactor we’d been putting off: moving the policy engine fully into the agent runtime. No new feature, no new rule type. Just a cleaner separation between where policies are evaluated and where policies are authored.
- blog#ai-governance
· Maurits Kaptein
Agents Don't Share a Language — And That's a Governance Problem
LangChain, CrewAI, the Microsoft Agent SDK, and Claude Code all describe agent behavior differently. Before you can govern agents, you need a common language for what they actually do.
- blog#ai-agents
· Maurits Kaptein
From Plain English to Enforceable Policy — Meet the Kyvvu Policy Generator
Describing what your agents shouldn’t do is easy. Formalizing that into deterministic, enforceable runtime policy is harder. We built an agent that bridges the gap.
- blog#ai-agents
· Maurits Kaptein
The Month AI Agents Went Rogue
March 2026 will be remembered as the month agentic AI incidents stopped being theoretical. Five real incidents. Real data exposed. Real systems compromised. Here is what happened — and why “just restrict the prompt” is not an answer.
- blog#ai-governance
· Maurits Kaptein
Policies on Paths: A Position Paper on Runtime Governance for AI Agents
We published a position paper on arXiv today. The core argument: the execution path is the right object for governance, not the individual action.
- blog#ai-agents
· Maurits Kaptein
Your Pipeline Might Be Governed. Your Agents Aren't.
An AI agent just sent a report to an external recipient it shouldn’t have contacted. Your CI/CD pipeline passed every check. Your deployment was clean. Your access control logs show nothing unusual — the agent had read access to the database and write access to email, both legitimately granted.
- blog#ai-agents
· Maurits Kaptein
Why Runtime Guardrails Are a Missing Layer in Enterprise AI
Many enterprises deploying AI agents seem to rely on careful prompting, thorough testing, and retrospective logging to ‘constrain’ their agents. This is however an insufficient model—and with the EU AI Act’s high-risk provisions taking effect in August 2026, it is also increasingly a regulatory liability.
- blog#ai-agents
· Maurits Kaptein
From Monitoring to Enforcement: The Three Layers of AI Agent Compliance
Most AI governance tools only solve one-third of the compliance problem. There are three distinct layers — monitoring, incident generation, and runtime intervention — and understanding the difference between them is not an academic exercise. For enterprises deploying agents in regulated environments, it is the difference between a defensible compliance posture and an optimistic bet.
- blog#ai-agents
· Maurits Kaptein
The Agent Dilemma: Power vs. Control — and Why Orchestration Is Now the Missing Layer
OpenClaw went viral, got absorbed by OpenAI, and accidentally started a fight with an insurance company. It perfectly illustrates the core tension in enterprise AI deployment: the agents powerful enough to be useful are the ones with the least governance. Here’s why orchestration is the missing layer.
- blog#ai-agents
· Maurits Kaptein
Mapping the EU AI Act to AI Agent Compliance
We’ve spent the past few months mapping EU AI Act requirements to what AI agents actually need to do at runtime. We’re sharing our findings because honestly, we think more people should be building agents—but in a way that’s compliant from day one.
- blog#ai-agents
· Maurits Kaptein
An Agentic World: What Happens When AI Agents Become Colleagues
Imagine a morning in 2030 where AI agents handle the noise, while you focus on strategy. This isn’t science fiction—it’s closer than you think. Based on the prologue of ‘AI Agents at Work.’