An AI agent stars in a cyberattack in Spain and carries out several phases autonomously

The Spanish Data Protection Agency (AEPD) has received for the first time a notification of a security breach in which an artificial intelligence agent would have autonomously participated in different phases of a cyberattack. The agency considers that the case represents a relevant change in the cybersecurity landscape, as the AI would not have limited itself to assisting a human attacker, but would have executed different actions on its own.

According to the information communicated by the affected company, the agent began by analyzing files in search of possible vulnerabilities and managed to log into the system. Once inside, it continued exploring possible access routes without constant human intervention. Finally, it would have found a vulnerability that allowed it to modify personal data and consult invoices. The AEPD still needs to analyze the information received and, for the moment, has not detailed how the credentials used to access the system were obtained.

From assistant to autonomous actor

The difference compared to a conventional AI system lies in its ability to receive an objective, divide it into different tasks, use tools, and adapt its next steps based on what it finds. In an attack scenario, this allows for the automation of processes that until now required much greater human intervention.

The AEPD considers that this type of incident marks a “qualitative change” in cybersecurity matters. An agent can analyze numerous files simultaneously, try different entry points, and modify its strategy during the process. The risk increases especially when it manages to access accounts, keys, or tokens with elevated permissions, as it could move between different services before security systems detect the activity.

A threat that is beginning to cease being theoretical

The Spanish case joins other international signals about the use of AI agents in offensive activities. Google has warned of operations in which artificial intelligence models have been used to solve problems during intrusions and automate different phases of attacks. In one of the cases cited, several agents managed to escalate privileges and compromise thousands of credentials in approximately six hours.

Given this scenario, the AEPD suggests that companies incorporate attacks executed through AI agents into their risk analyses and not consider them solely within traditional categories such as malware. Automation allows for increasing both the speed and scale of operations, while the National Cryptologic Center has also warned of this paradigm shift and has published recommendations to strengthen protection against these new threats.