Spanish authority: AI agent carried out data breach on its own

The Spanish data protection authority AEPD wrote on its own blog on September 14, 2026 that it has received the first report of a personal data breach in which the attack itself was carried out by an AI agent. An AI agent is a computer program built on a language model (the kind of technology behind chatbots like ChatGPT) that can be set to solve a task on its own, without a human guiding every single step.

According to the authority, the attack unfolded in several steps without a human steering it along the way. The agent first searched for security flaws in ordinary files on a website, then logged in with credentials it had found, continued on its own to look for weaknesses in the system itself and ended up altering personal data and viewing invoices.

The AI agent acted independently

What's new isn't the attack method itself, but that the agent acted so independently. It could receive an overall goal, break it down into smaller tasks, use digital tools, run code and continuously adjust its own approach, without step-by-step instructions from a human. AEPD writes that artificial intelligence doesn't create entirely new types of attacks, but makes already known attacks faster, larger in scale and better at adapting.

The authority has not disclosed which organization was hit, which language model the agent was built on, or which industry is involved. AEPD also stresses that a single case isn't enough to say anything general about how often this kind of thing happens. But according to the authority, the case shows that AI-assisted attacks have gone from something people imagined to something that actually happens.

Experts urge calm

The case has also been covered by the US outlet SecurityWeek, where security experts urge people to keep a cool head. Simon Phillips, CEO of the security firm CyberVerse, points out that there may be other explanations. Perhaps the agent got out of control during a test, perhaps it was manipulated into circumventing its own rules (known as a jailbreak), or perhaps it was an authorized security test that has been mistaken for a real attack.

AEPD refers to Articles 32 and 33 of the General Data Protection Regulation, which require companies to continuously update their risk assessments and factor AI-driven attacks into their security work. The authority's concrete advice is: know your own data, don't store more data than necessary, limit who has access to the systems, close known security holes, keep track of your vendors, and have a plan ready in case something goes wrong.

What it means for Danish companies

For Danish companies and ordinary AI users, there's no indication that anything similar has happened in Denmark. But an EU country has now for the first time formally registered a case in which an AI agent carried out an entire attack sequence on its own. That's relevant because many Danish companies are themselves in the process of adopting AI agents for tasks such as customer service and system access these months. The case shows that the same type of technology could in principle be misused by others to attack systems very quickly.