Spain’s data protection authority has logged what it considers a first: a breach in which the intruding AI agent operated without anyone directing it.
The AEPD’s account comes from the filing the affected company submitted. By that document, the agent swept ordinary files for weak points, got itself authenticated, then probed the application in search of a weakness. That weakness gave it a way in to alter personal records and touch invoice data.
Francisco Pérez Bes, the agency’s deputy director, urged caution. The facts are still the victim’s own version and need analysis, he said; that a model appears in the case does not mean either it or its maker’s systems were breached, and nothing indicates the tool was built for attacks.
Even so, he treated the report as meaningful. One case proves no pattern, he said, yet it shows AI-assisted attacks have stopped being hypothetical and now land on live personal-data operations.
The agency reads AI as an accelerant for methods defenders already know rather than a new class of attack, adding pace, breadth and adaptability, which leaves less time to spot and stop an intrusion. Spain’s National Cryptologic Center arrived at much the same judgment in its offensive AI guide, treating the technology as a capability already in routine use across active campaigns.
Both bodies steer toward the same discipline: map what you process, keep fewer records, narrow who can reach them, patch sooner, hold suppliers to account, and govern how agents are run and watched.
Pérez Bes framed the response as a review task rather than a purchase, saying security and data protection models need immediate reassessment now that agents have entered the offensive arena. Data protection officers and managers, he added, should brace for faster attacks while leaning on fundamentals that have not changed.
