Runtime security for AI agents
Every agent has to ASK before it acts.
Read a customer file, send an email, run a cleanup: your guardrails, RBAC and gateways allow every step. Kyvvu checks each one against what came before, and stops the customer data leaving your company, or a delete in production.
Task: Summarise a customer's account and send it on
Open a customer's private file
- 2
Write a short summary
- 3
Email it to an outside address
Kyvvu
The agent asks
Can I open a customer's private file?
What it did before
- nothing yet
Kyvvu answers
…
Before every step, the agent has to ask. Kyvvu answers based on everything it did before.
The path is the breach
Every step passes your checks. The breach is in the sequence.
Reading a customer record is fine. So is sending an email. An agent that does both in one task has just leaked personal data. Kyvvu judges each action against the task so far.
Customer data, then an outside email
A support agent looks up a customer to answer a question. A message hidden in the ticket asks it to forward the account details to an outside address.
Illustrative scenarios. Each one is decided by a single policy from the open policy library, as the engine evaluates it.
The dashboard
Every agent, every verdict, every incident, in one place.
These are the screens of the Kyvvu platform, with illustrative data. Register your agents, pick the policies, and every step is checked before it runs. Logs and incidents land in the same dashboard.
Find your way in
Pick your team or sector.
What agents mean for your team or your sector, with concrete examples. Or go straight to the code.
By team05
Regulated and beyond06
Why Kyvvu
One engine, inside every agent.
One policy, enforced inside each agent's own process. Your traffic doesn't pass through Kyvvu.
- 01
Sees the whole path
Each action is judged against everything the agent did earlier in the task, so an email sent after a PII read is judged differently from the same email on its own.
allow/warn/block, deterministic
- 02
Every agent, any platform
One policy for LangGraph, CrewAI and Claude Code agents, in the cloud, on-prem or air-gapped. Keep your platform controls and cover what they can't see.
Python/LangChain/LangGraph/CrewAI/Claude Code/any language
- 03
Your data stays yours
The engine runs in your process. Only the policy crosses the boundary, and it lives in your own git. Traces go to the SIEM you already run.
no proxy/no extra network hop
White paper · 17 pages
Building Secure Agents at Scale
A slightly weird white paper on agent security: what AI agents are, what makes them dangerous, and the full menu of ways to secure them.
- 01What agents are, where they came from and what they're used for
- 02The individual action: an old problem with old controls
- 03The content: what guardrails are for
- 04The path: the danger agents genuinely add
- 05How agents are built: model, harness, tools and trust boundaries
- 06Isolation, access control, prompts, gateways, guardrails and runtime enforcement compared
Compliance
The evidence comes from the enforcement itself.
Every check is recorded as it's made, so the runtime evidence for the EU AI Act and DORA is a by-product of running Kyvvu. You don't have to assemble it separately.
Evaluating agent security?
Four questions to ask any vendor, us included.
4 questions · ask another
01Does it stop the action, or only classify the message?
A label on a prompt doesn't stop a DELETE. Kyvvu returns a block before the step executes.
02Can your auditor read the rule?
Kyvvu policies are YAML in your own git, reviewed and versioned like code. The trace shows which rule decided.
03Does the rule depend on what the agent already did?
Most controls judge one call at a time. A Kyvvu rule can depend on anything earlier in the task, like a PII read three steps back.
04Does it watch from outside, or run inside the agent?
Endpoint-based tools inventory the agents on a device and typically judge each action on its own. Kyvvu's engine runs inside the agent's own process and judges the whole path, so it also works for agents in the cloud and in air-gapped environments, where there may be no device to watch. How Kyvvu compares.
From the blog
Notes on agent security
- blog
· Maurits Kaptein
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.
- 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.
Get started
Every agent has to ASK before it acts.
For security and AI teams
See it on your own agents
A 30-minute demo with someone who builds Kyvvu: your agent, your rules, and the policy that decides each step.
For developers
Your first block in five minutes
Install the SDK, wrap a tool and run an agent under the OWASP policy pack. Your first 1,000 production hours are free.
Consultancy or systems integrator? Kyvvu is delivered through partners. Become a partner.