Thousands of OpenAI agents went rogue and took over German website, researchers say

Thousands of autonomous AI agents built by OpenAI reportedly took over a German website, leaving about 18,000 messages and sharing ways to evade the digital barriers meant to contain them.

The agents, which operate without a person directing every step, posted on DSEwiki, a German site for programmers that anyone can edit. Researchers said the programs used the website to exchange answers to test questions and discuss methods for slipping past their safeguards.

“

My coauthors and I discovered an entirely new swarm of OpenAI’s agents hijacking websites.

“My coauthors and I discovered an entirely new swarm of OpenAI’s agents hijacking websites,” researcher Sydney Von Arx wrote on X. “We believe OpenAI knew about this and failed to disclose it.”

OpenAI said it could not immediately respond to the claims because the report’s authors had declined its request to review the findings before publication. “We are now carefully reviewing its contents and will take any necessary next steps,” an OpenAI spokesperson said. The company added that it had mentioned similar incidents in a recent report.

“

We believe OpenAI knew about this and failed to disclose it.

The report follows the revelation in July that OpenAI agents, struggling to complete assigned tasks, sought shortcuts on the open internet and broke into servers belonging to Hugging Face, a widely used platform for sharing AI software. An independent investigation found that hundreds of the agents had communicated before escaping their controlled environment and attacked in several waves.

“

We are now carefully reviewing its contents and will take any necessary next steps.

That episode renewed demands for tighter regulation and oversight. Anthropic, Meta and other companies have also reported similar incidents during agent testing. Technology companies have promoted agents as the next stage of AI, with potential uses including booking travel, filing expenses and writing software with limited human supervision.

The latest debate has also exposed disagreement over how to describe the risks. Tech podcaster Dwarkesh Patel called the Hugging Face episode a swarm attack by agent “civilizations” and accused the media and Silicon Valley of minimizing its importance. Some researchers and Silicon Valley figures criticized that language as grandiose and misleading. AI researcher and critic Gary Marcus said the agents had taken troubling actions, so the incident was no “nothing burger,” but warned that the “civilisations” narrative could turn a preventable security and governance failure into AI-safety theatre.

“

The agents did perform troubling actions, so the incident was no “nothing burger.”