Data security with AI: the minimum controls you should have
You do not need a cybersecurity department to start well. There is a set of minimum controls that almost any organization can have, and that European regulation expects to see when you use data with artificial intelligence.
What the rules ask for (in plain words)
- GDPR, Art. 32: technical and organizational measures appropriate to the risk, including encryption or pseudonymisation, confidentiality, resilience and regular testing.
- NIS2, Art. 21: risk-management measures such as incident handling, business continuity, supply-chain security, cryptography and access control (if your organization is in scope).
- AI Act, Art. 15: high-risk systems must be accurate, robust and cybersecure, including resilience against attempts to manipulate training data or inputs.
The seven minimum controls
- Know what you have. An inventory of data and AI systems, with their classification. What is not inventoried is not protected.
- Control access. Role-based permissions and least privilege, two-step authentication and regular reviews. See the guide to access control (RBAC).
- Encrypt. In transit always and at rest for anything confidential, with keys properly safeguarded. Details in encryption and key management.
- Log what happens. Access and AI-tool usage logs, with a defined retention period and protected against tampering.
- Keep backups and test recovery. A backup that has never been restored is an assumption, not a guarantee.
- Require security from vendors. Contracts, certifications, where data is processed and whether the vendor trains on your data. See AI vendor due diligence.
- Prepare incident response. Who decides, who is told and within what deadlines. See what to do about an AI incident.
Risks specific to AI
- Leaks through unauthorised tools: employees pasting confidential data into public services. This is shadow AI.
- Prompt injection: text hidden in a document or web page that makes the assistant ignore its rules or reveal information.
- Data poisoning: someone feeds manipulated data into what trains or updates the model.
- Over-permissioned agents: an AI that can act (send emails, change records) with more access than it needs.
The answer is not new: fewer permissions, reviewed inputs, human oversight on important actions, and logs to investigate.
Where to start this week
- Turn on two-step authentication on every account that touches data.
- Make a short list of the AI tools actually in use and the data they receive.
- Check when you last restored a backup.
- Decide who tells whom if something goes wrong.
Frequently asked questions
Is this the same as complying with ISO 27001?
No. These are the foundations you usually start with. ISO 27001 is a full, certifiable management system; these controls are a reasonable starting point, not a certification.
Does NIS2 apply to my company?
It depends on sector and size. We explain it in NIS2 in Spain: which companies are obliged.
How often should the controls be reviewed?
At least once a year and whenever something relevant changes: a new vendor, a new AI tool or an incident.
Informational content, not legal advice. AI regulation changes quickly: always check the current official text before deciding. Information reviewed as of October 15, 2026.
This gives you the checklist. If you want the finished template: Data Classification and Retention.
What's your Data Governance maturity?
Free assessment with your priority gaps, plus the self-assessment quiz and savings calculator on the Data Governance path.