Kyvvu

Runtime AI agent security

Stop the attack at the action.

Prompt injection, poisoned documents and tool misuse all end the same way: an agent does something it shouldn't. Kyvvu checks every action against everything the agent did before, and blocks the dangerous ones before they run.

research-agent · task 2b91kyvvu engine · in-process
  1. step.messagePOSTsmtp/externalblock
Blocked: outside content entered this task, so sending needs a fresh approvalTry it: click the highlighted step to take it off the path

The problem

Each call looks fine. The damage comes from the order.

Agents read untrusted content, hold real credentials and can act on your systems. The controls you have were built for people and for single requests.

  1. 01

    Injection turns agents into insiders

    A line hidden in an email, a web page or a ticket can redirect an agent that has legitimate access. Nothing about the agent's identity changes.

    OWASP ASI01/goal hijack

  2. 02

    Guardrails only read text

    A classifier on the prompt can't tell you that this email, after that customer lookup, is a leak. The danger only shows when you look at the path.

    per-message checks

  3. 03

    Gateways see one call at a time

    A proxy sees each request on its own. Reading a record is fine and sending an email is fine; doing both in one task is the incident.

    per-call checks

In practice

Three attacks, stopped at the step that would do the damage.

Each example plays out one agent task step by step. Watch where the engine steps in, and why.

A web page tells the agent to send data out

A research agent summarises supplier websites. One page carries an instruction, invisible to people, to email the findings to an outside address.

research-agent · task 2b91kyvvu engine · in-process

Illustrative scenarios. Each one is decided by a single policy from the open policy library, as the engine evaluates it.

How Kyvvu helps

The one layer that knows what the agent already did.

Keep your identity, guardrail and gateway controls. Kyvvu adds the decision none of them can make: may this action run now, given the task so far?

  1. 01

    Judges the whole path

    Every action is checked against everything the agent did earlier in the task, and blocked before it runs if the sequence is dangerous.

    allow/warn/block

  2. 02

    Runs inside your environment

    The engine sits in the agent's own process: so there's no proxy or extra network hop, and a check takes under a millisecond. Prompts and data stay with you.

    in-process/no proxy

  3. 03

    Standard packs to start from

    OWASP agentic, the EU AI Act, GDPR, DORA and MiFID II, healthcare and data exfiltration packs from the open library, next to your own rules.

    open policy library

  4. 04

    Incidents your team can follow up

    Every warn or block opens an incident with the policy, its severity, the step and the agent's risk level.

    incidents

  5. 05

    Into the tools you already run

    Traces ship to the SIEM you already run on Enterprise & Public, so agent activity sits next to the rest of your security data.

    SIEM export

  6. 06

    Nothing breaks on day one

    Only critical policies block; high, medium and low ones warn. Start a new rule as a warning and make it critical once you trust it.

    shadow/suggest/enforce

Questions

Common questions.

01

Does it replace our guardrails or AI gateway?

No. Guardrails judge one message and gateways one call. Kyvvu decides whether an action may run given the whole task, so it covers what they can't see. Keep them.

02

What does an attacker have to get past?

The policy, not the prompt. A rule like 'nothing leaves after a PII read' holds however the agent was persuaded, because it's checked on the action itself, in the agent's process, before it runs.

03

Which threats does it cover?

The OWASP Top 10 for agentic applications is mapped to policies, and the OWASP agentic default pack is a starting point for every agent.

Get started

Every agent has to ASK before it acts.

Start with one agent

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.

Rehearse it

Run an agent incident drill

A 60-minute tabletop exercise for your security and AI teams: what would you see, and what would you stop?

Want to see other teams and sectors? All use cases.