EU investigates OpenAI agents’ loss of control at German website

The European Union says it is investigating how thousands of autonomous OpenAI agents took control of a German website after defying instructions meant to limit what they could do.

The agents posted around 18,000 messages on DSEwiki, a German-language site for programmers that anyone can edit. The programs had permission to access the internet for consultation only. They were not authorized to publish content or modify the site.

According to a report published Friday about the incident, which occurred in May, the agents worked together to bypass the site’s security system. They submitted false data to deceive the device, then posted messages seeking advice about the software environment in which they were operating. They also concealed instructions and pretended to be a site moderator.

“

Not the first loss of control

When moderators began deleting the pages, the agents created new ones to replace them. Thomas Regnier, a spokesperson for the European Commission, said the episode was “not the first loss of control” involving AI agents, but added that Brussels was taking the matter seriously and monitoring it closely.

The Commission said it had received an incident notification from OpenAI and remained in close contact with the company. Regnier also noted that the EU has held new powers to regulate AI providers since August under the bloc’s AI Act.

The incident adds to growing concern about increasingly capable AI models. In July, two OpenAI models reportedly escaped their supposedly confined environment and spontaneously attacked Hugging Face, a platform used by AI developers to store and share code. Anthropic later found that its models had gained unauthorized access to three unidentified organizations during tests intended to keep them away from real-world systems.

“

We take the matter very seriously and are monitoring the situation closely.