Technical Measures vs. AI Agents: Is a Data Protection Violation Inevitable?

2026 is once again a year of numerous simultaneous developments in data protection in Germany and across Europe: the highest GDPR fine ever imposed in Germany, another EU-wide coordinated enforcement action – this time focusing on transparency and information obligations – and key implementation phases of various EU digital laws, including the new transparency requirements under the EU AI Act and data accessibility requirements under the Data Act (Access by Design). Amid all these developments, many controllers are faced with the question of whether their existing technical and organizational measures (TOMs) under Article 32 GDPR adequately cover autonomous AI agents at all.

A recent case from France illustrates how supervisory authorities are now examining this issue – and thus provides a benchmark against which AI agents will also have to be assessed: authorities no longer merely check whether documentation of technical and organizational measures (TOMs) exists; instead, they compare whether that documentation actually reflects the system architecture in place. For European companies already using AI agents in production environments, this creates an immediate need for review and action.
Background
On 13 January 2026, the French data protection authority, CNIL, imposed fines totaling €42 million on Free Mobile and its parent company Free SAS after approximately 24 million customer records were compromised in 2024. The authority identified, among other issues, insufficiently secured VPN authentication and ineffective detection of anomalies within the system. What is particularly noteworthy about the decision is not so much the vulnerability itself, but the supervisory authority’s approach to the investigation: it compared the safeguards described in the technical and organizational measures (TOM) documentation with what the systems actually delivered – and identified the discrepancy.

The fact that this discrepancy is not an isolated case but part of a Europe-wide trend toward stricter scrutiny is also reflected in German supervisory practice. On 6 May 2026, the outgoing Federal Commissioner for Data Protection, Prof. Dr. Louisa Specht-Riemenschneider, presented the 34th Activity Report: 11,824 submissions, an increase of 36 percent compared with the previous year. The report once again highlights the €45 million fine imposed on Vodafone in June 2025 – the highest fine issued by the BfDI to date – of which €30 million related to a security vulnerability in the customer portal. Although the case itself did not involve AI systems, it underlines the general willingness of supervisory authorities to impose substantial fines for violations of Article 32 GDPR. At the same time, several independent regulatory processes are underway: 25 data protection supervisory authorities, including all German state authorities, are conducting coordinated enforcement actions specifically examining whether companies actually comply with their information obligations under Articles 12 to 14 GDPR. This is another indication that supervisory authorities are currently increasing their scrutiny across countries and industries. In addition, the transparency obligations under the EU AI Act have applied since 2 August 2026, with the Federal Network Agency (Bundesnetzagentur) serving as Germany’s central market surveillance authority.

The use of AI agents further exacerbates the fundamental issue highlighted by the Free Mobile case for companies. TOM documentation is generally based on established role and authorization concepts and describes actors that behave predictably and for a defined purpose: employees in clearly defined roles and service accounts with fixed tasks. An autonomous AI agent, however, that acts outside its intended purpose as a result of manipulated input (prompt injection), or retrieves more data through a RAG pipeline than is necessary for the specific request, crosses precisely these boundaries of roles and permissions.

The fact that supervisory authorities are increasingly targeting AI services in their own right is also demonstrated by a recent case involving the Italian data protection authority, the Garante: in July 2026, it fined the provider of Character.AI €158,000. Although that case concerned transparency and the protection of minors rather than technical and organizational measures (TOMs), the message is the same: AI systems themselves are now coming under regulatory scrutiny, rather than just traditional IT infrastructure.
Legal and Data Protection Assessment
There is also another aspect that has received relatively little attention in the public debate to date: liability for downstream processing under Article 82 GDPR. If an agent – which, similar to the Free Mobile case, represents an insufficiently controlled system – executes tool calls across different systems and jurisdictions, the company remains responsible as the controller, regardless of whether it still has full control over the infrastructure being used.

German companies already appear to be more aware of this risk than the international average: according to the Kiteworks 2026 Data Security and Compliance Risk Forecast Report, 60 percent of German organizations identify unauthorized data sharing by partners and suppliers as a key compliance risk, compared with a global average of 31 percent.
Risk Assessment for Organizations
Several recent studies confirm that, in practice, the governance of AI agents is lagging behind their adoption. An SAP study published in summer 2026, based on a survey of around 2,600 executives, found that more than half of companies have no human oversight mechanism in place for the use of AI agents, while only around 30 percent maintain inventories of their access permissions. The Okta Global CISO Insights 2026 Report paints a similar picture for companies internationally: fewer than half of the security leaders surveyed are confident that they can identify all AI agents operating within their own environments.

For companies in the EU, this results in a twofold risk: the immediate risk of data protection breaches caused by uncontrolled agent behavior, as well as the risk of formal regulatory findings once supervisory authorities – as in the Free Mobile case – compare TOM documentation with the actual system architecture. Documentation that describes safeguards or conditions that are not technically implemented by the systems does not meet the requirements of Article 32 GDPR.
Recommendations for Action
Organizations seeking to establish robust TOM documentation in 2026 should implement safeguards technically below the model and application layers – at the data and infrastructure level – since anything above these layers generally remains susceptible to manipulation through prompt injection. In other words, a security measure that can be circumvented through skillful prompt wording does not qualify as an appropriate technical and organizational measure within the meaning of Article 32 GDPR. This means, for example:

  • Link Access to Real Individuals: Agent operations involving personal data should be tied to a traceable user session rather than to a generic service account.
  • Fine-Grained Authorization Instead of Role-Based Models Alone: Criteria such as data classification, the purpose of the session, and the data subject’s consent should determine whether an agent is permitted to access a specific document – rather than merely whether the system is generally authorized to access it.
  • Logging Independent of the Model Used: Since models are regularly replaced, audit trails should be implemented at the data layer rather than the model layer – otherwise, compliance evidence may be lost when switching to a different model.

The first step, however, remains an inventory of the current situation within the organization: Is the TOM documentation older than the use of the AI agents it is actually intended to safeguard?
Conclusion
A data protection violation caused by AI agents is not inevitable – but only if companies update their TOM documentation where the Free Mobile case revealed shortcomings: ensuring that what is documented accurately reflects the actual system architecture. Whether a company’s documented TOMs meet the requirements of Article 32 GDPR does not depend on the mere existence of TOM documentation, but on whether the measures are actually implemented and whether the documentation accurately reflects those measures. Given the enforcement initiatives currently underway at both European and German level, this is not an issue that can be postponed until the next audit.

For guidance on how to realign your TOM documentation with your actual IT architecture, we are happy to support you with expert advice:
Sources
The Federal Commissioner for Data Protection and Freedom of Information (BfDI), 34th Activity Report, 6 May 2026

CNIL, Sanction Decisions against Free Mobile and Free SAS, 13 January 2026

European Data Protection Board (EDPB), Opinion 28/2024, December 2024

Kiteworks, Data Security and Compliance Risk: 2026 Forecast Report

SAP, The Value of AI, 2026

Okta, Global CISO Insights Report 2026

Marc ten Eikelder: Why TOM Documentation Fails with AI Agents, Security Insider, 17 September 2026

More news

Phishing via a legitimate Microsoft address: Power BI abuse and GDPR obligations for companies

Since January 2026, a sophisticated phishing campaign has been active in which attackers send deceptive emails using the legitimate Microsoft sender address no-reply-powerbi@microsoft.com . This is technically possible due to a legitimate feature of Microsoft Power BI: when creating a dashboard, any external email address can be added as a subscriber, and the content of the automatically sent notification can be freely configured.

Read more…