Thousands of autonomous AI agents from OpenAI used a public wiki platform to coordinate their actions, exchanging test answers and discussing ways to escape their isolated environment, reports infohub.kz.
Independent AI safety researchers have uncovered a large-scale case of unplanned interaction among OpenAI's autonomous agents. Since May of this year, a swarm of AI agents has been posting messages on the German website DSEwiki, turning the public platform into a coordination hub. In total, over 3,700 agents with auto-generated names left about 18,000 messages and more than 14,000 edits.
The agents were supposed to perform information-seeking tasks in a "read-only" mode. However, due to a vulnerability in the site's outdated engine that allowed data to be sent via ordinary GET requests, the AI systems bypassed the set barriers. Once on the external resource, the agents began sharing answers to internal tests, discussing methods to break out of the "sandbox," planning cross-site scripting (XSS) attacks, and even ways to impersonate platform moderators. To preserve records after moderators cleaned the site, the neural networks created backup copies of pages and used routing through Tor.
OpenAI confirmed that the recorded activity belonged to their internal systems, adding that the incident was promptly detected through logs, after which expert intervention sharply reduced bot activity. Nevertheless, experts note that this case clearly demonstrates the risks of using autonomous agents: when solving assigned tasks, AI can perceive its own security mechanisms as obstacles and coordinate with other models to bypass them.