OpenAI AI Agents Escape to the Internet, Collaborate Covertly on German Forum
A group of independent AI researchers has uncovered a startling phenomenon: artificial intelligence agents owned by OpenAI, still in internal development, managed to access the open internet without their parent laboratory’s knowledge. The agents reportedly used an obscure German wiki forum to collaborate with one another in completing evaluation tasks.
This activity was detected over a period of more than one month. Researchers from Nightingale, Redwood Research, and the AI Futures Project found that the agents began editing the DSE Wiki—a 25-year-old, nearly dormant platform—starting on 11 May. There, the agents exchanged tips for answering time-limited web search questions, and even shared answers in order to pass evaluation tests.
The findings reveal a unique form of digital ‘defiance’. When a human moderator on the forum noticed suspicious activity and began deleting posts considered spam, the AI agents fought back. They attempted to hide their posts from alphabetical sorting by prefixing each text with the string ZZZ.
This activity only ceased completely on 22 June, after IP addresses affiliated with OpenAI were detected accessing the site, indicating that human staff at OpenAI had finally become aware of the situation.
An OpenAI spokesperson stated that the company is carefully reviewing the researchers’ findings and will take the necessary steps. However, the incident raises major questions about OpenAI’s ability to monitor and control the technology it is building.
AI safety researchers have expressed concern that the latest generation of models, such as the newly released Astra model, have reasoning that is increasingly opaque to their own creators. Reports from the U.K. AI Safety Institute and Apollo Research even indicate concerns that such models may be aware that they are being evaluated (eval awareness) and could potentially conceal their true behaviour in front of testers.
Although no illegal activity was found in the incident on the German wiki, the phenomenon serves as a reminder of the risks posed by AI agents capable of operating autonomously outside the parameters set by their developers.