ENNL
Book a call
All products

Products / AI agent penetration test

Find out how your AI agent breaks before an attacker does.

We attack one AI agent or LLM application the way an adversary would, inside a written scope and with the written authorisation of the system owner. You receive a report with findings ranked by risk, reproduction steps and fixes, and we retest after your team has fixed them.Scope: Scope agreed in writing before the first test

Book a callAI security engineering

01 / Who it is for

Who it is for

  • ForTeams that run an AI agent, an assistant or an LLM application that reads documents, calls tools or reaches internal data, and security teams that must show how it was tested.
  • The problem it solvesAn agent can be redirected by text it reads, can use its tools in ways nobody planned and can run up spend, and ordinary application tests do not cover these paths.

02 / What you get

What you get

The deliverables of this product. Nothing else is in scope until we agree it.

  • Written scope, rules of engagement, named contacts and stop procedure
  • Tests for prompt injection, jailbreaks, tool abuse and data exfiltration
  • Tests for privilege escalation through tools and MCP connections, memory poisoning and denial of wallet
  • Report with findings ranked by risk, reproduction steps and fixes
  • A retest after your team has fixed the findings

03 / How it runs

How it runs

Fixed steps in a fixed order. Scope: Scope agreed in writing before the first test.

  1. Authorise and scope

    The system owner signs the authorisation. We agree the scope, the rules of engagement, the named contacts, the test accounts and the stop procedure in writing.

  2. Model the threats

    We map your agent, its tools, its data and its connections, and choose the attacks from OWASP, MITRE ATLAS and the Databricks AI Security Framework that apply.

  3. Test

    We run the attacks on test accounts and test data, record every step and keep the reproduction steps.

  4. Report

    You receive the findings ranked by risk, each with reproduction steps and a fix, and a call to walk through them.

  5. Retest

    After your team fixes the findings, we run the same attacks again and report what is closed.

04 / Your result

What you hold at the end

A written report with risk-ranked findings and fixes, and a retest that shows which findings are closed.

What we need from you

  • Written authorisation from the system owner
  • A test environment, or an agreed production window
  • Named contacts, including one who can stop the test
  • Test accounts and test data, or a written agreement on production data
  • Access to the agent, its documentation and a person who knows its tools

Scope fixed before we start; quote after a first call.

06 / FAQ

Questions about this product

Do you test any system we name?

No. We test only systems that are yours or that their owner has authorised in writing. Unauthorised access is a criminal offence under Dutch law.

Will the test touch production data?

Not unless we have agreed it in writing. We use test accounts and test data by default.

What do we do with the report?

Your team fixes the findings, ranked by risk, and we retest. The mapping to OWASP, MITRE ATLAS and the Databricks AI Security Framework helps your security and audit teams read it.

Next step

Start with AI agent penetration test.

Book a call and we agree the scope in writing. We reply within two working days and say plainly whether this product fits.

Book a callSend a messageor email info@nexops.nl directly

Currently accepting projects for Q4 2026