Data breaches are no longer solely the result of technical flaws in information systems. Since the widespread adoption of generative AI tools in businesses, a new exposure channel has emerged: prompts, attachments sent to models, and generated responses all represent potential exit points for confidential information. Protecting sensitive data now requires mapping these new perimeters as well as the classic vectors.
Generative AI and Sensitive Data: An Underestimated Leakage Perimeter
The CNIL finalized its recommendations on AI in July 2025, highlighting a specific risk: sensitive data can be exposed in training datasets, in documents submitted by users, and in the responses produced by the model. The deployment mode (on-premises or cloud), the conditions for reusing inputs by the provider, and transfers outside the European Union are three criteria to evaluate before any production rollout.
This risk is not theoretical. An employee who pastes a contract excerpt into a cloud chatbot potentially transmits this content to a server outside the EU, without the data controller being aware. The question is not to ban the tool, but to regulate what is injected into it.
For those looking to secure their sensitive data with Jeune et Actif, the process begins with an audit of outgoing data flows, including those that pass through artificial intelligence services.
Obligation to Control AI: What the AI Act Changes in Practice
Since February 2, 2025, Article 4 of the AI Act imposes an obligation to control AI on providers and deployers. This provision goes beyond the usual scope of cybersecurity: it requires individuals using AI systems to have a sufficient level of competence to understand the associated risks.
In practice, this means that traditional cybersecurity awareness training is no longer sufficient. Specific rules regarding prompts, attachments, and confidential data sent to models must be added. An employee trained in strong passwords but unaware of the risks associated with an AI assistant remains an exposure vector.
Field feedback varies on the maturity of companies in facing this obligation. Some have integrated AI usage charters as early as 2024, while others have not yet identified which generative AI tools are actually used by their teams. The gap between the two creates a zone of regulatory as well as technical risk.
Ransomware and Exfiltration: Why Backups Alone No Longer Provide Protection
Ransomware has changed in nature. ENISA still identifies this threat as having the most significant impact on organizations. The difference from attacks a few years ago lies in the systematic combination of three levers: data encryption, prior theft, and extortion under the threat of publication.
In the face of this triptych, restoring a backup only addresses one-third of the problem. If data has been exfiltrated before encryption, disclosure remains possible even after complete system restoration. Protection must therefore occur upstream, across multiple simultaneous axes:
- Segment access so that a compromised account does not provide access to the entire informational asset
- Monitor abnormal behaviors on the network, particularly unusual volume transfers to external destinations
- Regularly test incident response procedures, not just backups, but the entire chain of detection, containment, and communication
Limiting exfiltration has become as much a priority as ensuring restoration. This paradigm shift involves different investments: less backup storage, more real-time supervision.
Cloud and Data Sovereignty: Choosing Your Deployment Model
The choice between public cloud, private cloud, and on-premises infrastructure directly affects the level of control over sensitive data. The CNIL recommends evaluating this criterion upfront for any project involving AI, but the reasoning applies to all critical data processing.
Transfers outside the European Union remain a major friction point. A cloud service hosted in Europe but operated by a provider subject to extraterritorial legislation (such as the U.S. Cloud Act) can expose data to access requests from third-party jurisdictions. Available data does not allow for a conclusion that standard contractual clauses provide sufficient protection in all scenarios.
For companies handling particularly sensitive data (health, defense, judicial data), deployment on-premises or on a SecNumCloud-qualified cloud by ANSSI represents a more protective option, albeit at the cost of greater operational complexity.
Building a Data Protection Policy that Integrates These New Risks
Classifying data remains the foundation of any security strategy. Not all information deserves the same level of protection, and applying maximum measures everywhere amounts to prioritizing nothing. Classification allows for the allocation of monitoring and encryption resources where the impact of a breach would be most severe.
Beyond classification, three dimensions deserve to be integrated into existing policies:
- An explicit AI component, defining which types of data can be subjected to generative AI tools and under what deployment conditions
- An exfiltration component, with alert thresholds on outgoing data volumes and network segmentation rules adapted to asset criticality
- A regulatory compliance component covering both GDPR and the AI Act, two now complementary frameworks for any organization processing personal data with automated tools
The challenge lies in articulating these components. Legal, IT, and business teams must work together on concrete scenarios rather than generic charters. An annual penetration test does not cover the risk associated with unregulated daily use of a chatbot by the sales department.
Protecting sensitive data is no longer limited to a firewall and a weekly backup. The perimeter to defend now includes every interaction with an AI system, every cloud flow, and every abnormal user behavior. Organizations that do not update their risk mapping in the coming months will expose themselves to both regulatory sanctions and operational losses.
