AI Security in Europe : ROT Data Emerges as a Central Challenge

AI Security in Europe : ROT Data Emerges as a Central Challenge
3 min read

As the European Union prepares to enforce its Artificial Intelligence Act, the quality of the data feeding AI systems is attracting less attention than their compliance. The authors of the analysis argue that a model can pass conclusive tests and meet regulatory requirements without its technical reliability being guaranteed.

Uncertainties surrounding the deadlines of the European AI law, as well as the new requirements expected under the EU’s cybersecurity and artificial intelligence action plan, are a major concern for European companies’ Chief Information Security Officers (CISOs). To be considered ‘trustworthy’, an AI system must successfully pass tests in sandboxes, i.e., isolated testing environments, and meet specific regulatory standards. However, official compliance and operational security are not necessarily synonymous.

Data is the most often overlooked link. If the data provided to models is redundant, outdated, or trivial—so-called ROT data—the controls are no longer valid.

Still Incomplete Standards

European AI regulation has continued to evolve in recent months. As the technical standards needed to assess compliance are not yet finalized, the main compliance deadline applicable to high-risk AI systems has been pushed back from August 2026 to December 2027.

To compensate for the lack of assessment frameworks, the European Union published its action plan several weeks after announcing this postponement. This plan adds no legal constraints. Instead, it provides for protected testing environments, guidelines, and assessment tools to be used prior to any commercial launch, with the aim of ensuring cutting-edge AIs are deployed safely and remain compliant with the AI Act and the NIS2 Directive.

This framework fosters the development of secure models, according to the analysis, but it only covers part of the chain. To obtain compliance certificates, security teams will need to examine code in detail and gather supporting documentation. However, a secure and compliant AI tool can still produce irregularities if misused.

A human error is, in most cases, detected before it escalates. Yet non-human identities and autonomous agents are 82 times more numerous than human employees in medium-sized companies. With AI agents, errors spread very quickly, and defective information can corrupt the best-protected model without triggering an alarm.

The Hidden Cost of Data

Companies scrutinize models carefully, but the information that feeds them receives less consideration. When the issue of data is addressed, it is primarily from a security angle and rarely from a quality perspective.

Approximately 80% of data stored by companies is in unstructured formats and includes a large amount of ROT files, accumulated over several decades in obsolete cloud infrastructures. Feeding AI agents with this data can corrupt them and lead to decisions based on skewed analyses, to the point of causing non-compliance situations, even when tests conducted in sandboxes prove conclusive.

Managing this parasitic, often duplicated, data also generates costs that are difficult to identify in organizational accounting. Computing and cloud resources are mobilized to process irrelevant information. The capital expenditures (CapEx) of major technology groups dedicated to artificial intelligence have surpassed the $650 billion mark, but European texts, the analysis notes, still ignore the cost of the data feeding these systems.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x