The Spanish Agency for Data Protection (AEPD) has received the first notification of a personal data breach in which the attack would have been executed using an artificial intelligence agent. The system, which used a well-known language model, managed to access an application and, once inside, sought ways to autonomously find vulnerabilities until it succeeded in modifying personal data and accessing invoices.
The Agency itself urges caution regarding what happened. The available information comes from the notification submitted by the affected organization and still needs to undergo the corresponding analysis, so the AEPD does not definitively confirm all the details of the incident.
It also does not identify in the information released which language model was used or which organization suffered the breach. Furthermore, it emphasizes that the use of a certain model does not mean that the AI system itself or the infrastructure of the company providing it has been compromised, nor that the tool was designed to carry out malicious activities.
The agent sought vulnerabilities by itself
According to the information communicated to the AEPD, the attacking agent began by performing a search for vulnerabilities in generic files and successfully logged into the system.
Once access was obtained, it began to autonomously search for vulnerabilities in the application. Ultimately, it managed to exploit them, allowing it to alter personal data and access invoices.
For Data Protection, the main novelty is that a third party would have used an AI agent to successfully chain different phases of a cyberattack, instead of limiting the use of artificial intelligence to assist a person in specific tasks.
The Agency warns that a single notification does not yet allow for speaking of a statistical trend, but it does consider the case a relevant signal that AI-supported attacks are beginning to materialize in incidents that affect real personal data processing.
The leap from AI as an assistant to AI that acts
The use of artificial intelligence for malicious activities does not begin with this incident. Generative models can already be used to create phishing messages, translate fraudulent campaigns, impersonate identities, analyze code, or facilitate the search for vulnerabilities.
The difference that the AEPD highlights is in the emergence of the so-called AI agents. These systems can receive a goal and, based on it, plan intermediate tasks, use different tools, execute code, consult information, and interpret the results to subsequently adapt their actions autonomously.
Therefore, artificial intelligence does not necessarily generate new classes of threats, but it can increase the speed, scale, and adaptability of attack techniques that already existed. The consequence is a reduction in the time margin that organizations have to detect and stop an intrusion.
The National Cryptological Center had already warned of the risk
The warning coincides with the conclusions gathered by the National Cryptological Center (CCN) in its guide to best practices against the offensive AI model.
The agency considers that offensive artificial intelligence is evolving towards an operational capability integrated into real campaigns. Among its recommendations are to accelerate vulnerability management, strengthen identity protection, control the supply chain, and establish appropriate governance for the use of agents.
The incident reported to the AEPD now represents a specific case in the field of data protection in Spain in which an AI agent would have participated directly in the different phases of the attack.
Companies will have to review how they respond to cyberattacks
The AEPD believes that the case forces to rethink the risk analyses of organizations. In its view, it is no longer sufficient to generically consider threats such as malware, phishing, or unauthorized access, but attacks assisted or executed through artificial intelligence must be expressly incorporated.
It also points to reaction times. Procedures designed to address manually developed attacks may be insufficient if an agent is capable of simultaneously analyzing different assets, testing multiple entry points, and quickly modifying its behavior.
Another critical point is credentials. If an agent gains access to a account, an API key, or a token with excessive permissions, it can move between different services at machine speed before the organization identifies the anomalous behavior.
The Agency concludes that data protection cannot rely exclusively on manual intervention. Human supervision remains necessary, but it must be complemented with sufficiently fast detection, containment, and response systems to address automated attacks using AI.