# AI agent safety scan: is your AI agent secure under the GDPR and the AI Act?

> Is your AI agent secure and compliant? An opinion for each agent, with my name under it

Page: https://www.praxikon.com/en/diensten/agent-veiligheidsscan
Updated: 2026-11-06
Provider: Praxikon, a trade name of Embed AI (Dutch Chamber of Commerce 90283597), Rivium Westlaan 46, 2909 LD Capelle aan den IJssel, the Netherlands. The work is done by Zahed Ashkara, freelance AI and privacy legal adviser, exclusively as a freelancer.

For each AI agent I map what it may read, send and change, and test what the GDPR, the AI Act and, if it applies to you, the Dutch Cybersecurity Act ask of it: the dangerous triangle, human approval, logging and a stop switch, disclosing that it is AI, and the agreements with the vendor. For each agent you receive an agent passport, an incident reporting guide and a signed opinion with the three measures that remove the most risk.

## Price and terms

Fixed price per agent, up front, proposal within three working days of the call.

## Who it is for

- Organisations that use or build AI agents: a customer service agent, an assistant with access to email or case files, an agent in Microsoft Copilot Studio or a home-built one
- Municipalities and other public bodies that want to deploy an agent and must be able to explain it to their board and council
- DPOs, privacy officers and CISOs who need evidence for the board, procurement, the auditor or the insurer

## Not for

- A technical penetration test: that is for your security partner. I supply the test questions and assess the outcome legally
- A guarantee that an agent cannot be misled: no such guarantee exists. The scan makes the risks and the measures visible and demonstrable

## What you get

- For each agent an entry in the agent register: task, owner, vendor, permissions, data and connections
- The dangerous triangle test: which leg can go, and which prior approval is needed in addition, including for fetching web pages
- Human approval, logging and a stop switch tested, with Article 22 GDPR where the agent decides about people
- Disclosure that it is AI (Article 50 AI Act) and the agreements with the vendor (Article 28 GDPR) checked
- An incident reporting guide: which incident you report where and within which deadline, under the GDPR, the Dutch Cybersecurity Act if it applies to you, and the AI Act
- An agent passport for each agent: one page for the board, procurement or the insurer
- A signed opinion with the three measures that remove the most risk

## How it works

1. **A 30-minute call**: You explain which agents you use or build and what they do. I tell you honestly whether the scan fits.
2. **Proposal within three working days**: A fixed price per agent and a schedule. No open end.
3. **Inventory per agent**: With your administrator I go through the permissions, data and connections of each agent, usually in one session.
4. **Delivery, signed**: Agent passport, incident guide and the signed opinion, followed by thirty days for questions.

## Lead time

2 to 3 weeks

## The law behind it

- art. 32 AVG: https://www.praxikon.com/en/avg/artikel/32
- art. 25 AVG: https://www.praxikon.com/en/avg/artikel/25
- art. 33 AVG: https://www.praxikon.com/en/avg/artikel/33
- art. 50 AI Act: https://www.praxikon.com/en/ai-act/artikel/50
- art. 26 AI Act: https://www.praxikon.com/en/ai-act/artikel/26

## Questions

### What does an AI agent safety scan cost?

A fixed price per agent, which you see up front in the proposal. After a 30-minute call you receive the proposal within three working days; you do not pay for the call. The contracting party is Embed AI (Dutch Chamber of Commerce 90283597).

### What is the dangerous triangle?

An agent can be hijacked when it can do three things at once: read content from outside (email, websites, documents), see confidential data and send something out by itself. A hidden instruction in an email or web page can then make it send data away. The term comes from Simon Willison (the lethal trifecta). The fix is to remove one leg. Human approval up front is an additional measure and must then cover every outward channel, including fetching a web page, because an agent can put data in a web address.

### Is this a penetration test?

No. A penetration test tries to break in technically and belongs with your security partner. The scan looks at what the agent may do and what the law asks of that, and supplies the test questions a penetration test can include.

### Does the AI Act apply to our agent?

An agent is usually an AI system; the European Commission calls that a provisional reading in its frequently asked questions. If the agent interacts with people, the provider must ensure they are informed it is AI at the latest at the first interaction, unless that is obvious (Article 50(1) and (5) AI Act). Whether the agent is a high-risk system depends on the task, for example recruitment or creditworthiness (Annex III), not on the technology.

### The agent comes from a vendor. What is left for us to do?

You deploy the agent and remain responsible for your data and your choices: which permissions you give it, which connections and who approves. If the vendor processes personal data on your behalf, a data processing agreement is required (Article 28 GDPR). If the vendor may use your data to train its model, it becomes a controller for that purpose itself and you need a legal basis to provide that data.

### What do we have to report when an agent gets something wrong?

That depends on what happens. If an agent leaks personal data, you notify the Dutch Data Protection Authority without undue delay and where feasible within 72 hours, unless the breach is unlikely to result in a risk; recording it internally is always required (Article 33 GDPR). If your organisation falls under the Dutch Cybersecurity Act, you report a significant incident without delay and in any event within 24 hours as an early warning, within 72 hours as a notification and within a month with a final report, to the CSIRT (as a rule the NCSC) and the competent authority, a minister per sector (Articles 25 to 29 of that Act). For a serious incident with a high-risk system the provider reports to the market surveillance authority immediately and at the latest within 15 days, within 10 days in the event of a death and within 2 days for a widespread infringement or a disruption of critical infrastructure (Article 73 AI Act); as a deployer you immediately inform first the provider and then the importer or distributor and the market surveillance authority (Article 26(5) AI Act). For high-risk systems under Annex III the AI Act duties apply from 2 December 2027. If the system was already on the market or in use before that date, they apply only after a significant change in its design; systems intended for public authorities must comply by 2 August 2030 at the latest (Article 111(2) AI Act). The GDPR and the Dutch Cybersecurity Act already apply. The incident guide sets this out for each agent.

### Who does the work?

Zahed Ashkara, freelance AI and privacy legal adviser, does the work himself and puts his name under the deliverable. He works exclusively as a freelancer, not through secondment or employment.

### When can the work start?

We agree the start date in the 30-minute call. Send the question or assignment description in advance, and you will know in that call whether and when it fits.

## Contact

- Book a 30-minute call: https://cal.com/zahed-ashkara/30min-en
- Email: info@embedai.nl

_This text is the Markdown version of the service page; where they differ, the page prevails._