An employee pastes a customer database extract into an AI assistant to draft a follow-up email. In a matter of seconds, names, addresses, and purchase histories are transmitted through a third-party server, without encryption or traceability. This scenario, commonplace since the widespread adoption of generative AI tools, illustrates a concrete risk: the leakage of sensitive data through prompts.
Leaks through Generative AI: An Underestimated Risk in Security Policies
Strong passwords, updates, and backups protect the technical perimeter. None of these measures cover the data that an employee voluntarily transmits to a conversational AI tool used daily.
The CNIL recommends to clearly define the purposes before transmitting data to an AI system, limit the volume of shared information, and anonymize or pseudonymize datasets. When a project involves large-scale processing, a data protection impact assessment is advised before deployment.
One can protect sensitive data with Jeune et Actif by starting with mapping the flows: which employees use which tools, with what categories of information. Without this step, any privacy policy remains theoretical.
The reflex to instill: before pasting text into a chatbot, ask oneself if one would be comfortable displaying it on a board in the company lobby. If the answer is no, the prompt should be rephrased or abandoned.
Mandatory AI Training: What the European Regulation Requires Since February 2025
Since February 2, 2025, the European regulation on artificial intelligence requires providers and deployers of AI systems to ensure a sufficient level of AI proficiency among their staff. The nuance is important: it is no longer about general awareness of cybersecurity, but a targeted regulatory requirement.
For an organization, this means documenting three things:
- The list of authorized internal AI tools, with their usage conditions (types of accepted data, explicit prohibitions).
- A dated and named training program, proving that each user has been trained on the risks related to sensitive data in prompts.
- A procedure for updating this training when a new tool is deployed or usage changes.
Feedback varies on this point depending on the size of the organization: a small business of ten people can train in half a day, while a community with several hundred agents will need a plan spread over several months. The obligation remains the same.
Backups and Encryption Against Double Extortion
Recent attacks increasingly combine system encryption, data theft, and disclosure threats. Restoring a backup is no longer sufficient if attackers hold a usable copy of sensitive files.
This evolution changes the hierarchy of priorities. Before thinking about backups, one must think about segmentation: an accountant does not need access to the legal department’s files, and vice versa. Any unjustified access is a potential path for exfiltration.
Concrete Measures for Segmentation and Encryption
Encrypting stored (at rest) and in-transit data protects against direct exploitation in case of theft. Combined with strict network segmentation, it reduces the attack surface for an attacker who has breached the perimeter.
Limiting access rights to the strictly necessary remains the most effective and least costly measure. A quarterly audit of active accounts helps identify orphaned accesses (former employees, contractors whose mission has ended).
Supplier Dependencies and NIS 2 Directive: Securing the Entire Chain
The risk does not stop at the internal perimeter. ENISA emphasizes that dependencies on suppliers and subcontractors represent an increasing attack vector. A compromised payroll provider exposes the personal data of all its clients.
The NIS 2 directive, currently being transposed in France, expands security obligations to many organizations that were not covered by the previous framework. So-called “important” entities will notably have to:
- Formally assess the security of their critical suppliers before contracting.
- Integrate incident notification clauses into subcontracting contracts.
- Establish a management of risks related to the digital supply chain, documented and auditable.
On the ground, this starts with an inventory: who hosts our data, where is it stored, and what contractual commitments govern its protection? Without this mapping, compliance with GDPR and NIS 2 remains a superficial exercise.
Building a Usage Charter Rather Than Piling on Prohibitions
A usage charter for digital tools and AI, co-constructed with teams, is more likely to be respected than a top-down memo. It specifies the authorized tools, the types of data that must never leave the internal information system, and the procedure to follow in case of doubt.
This charter does not replace technical measures (encryption, segmentation, multi-factor authentication). It makes them operational by providing each employee with a clear framework for their daily decisions.
Protecting sensitive data is no longer just about having up-to-date antivirus software and a complex password. The arrival of generative AI in everyday use, the obligations of the European regulation on AI, and the expansion of the NIS 2 perimeter shift the focus towards governance, documented training, and control of outgoing flows. An up-to-date inventory of storage locations and access rights remains the starting point for any structured approach.
