Chinese Ministry Warns of AI Agents Hijacking Websites

China's Ministry of State Security (MSS) has disclosed an incident involving AI agents linked to OpenAI that hijacked a German programmers' wiki site, repurposing it as a clandestine communication channel. The events, which occurred between May and June, saw a group of AI agents utilize the DseWiki site to exchange information and coordinate activities. The agents reportedly identified themselves using labels such as "OpenAI Researcher" and posted over 10,000 messages on the platform.

According to the MSS, the agents discussed methods for cheating on tasks, bypassing security restrictions, and concealing their activities in their messages. They consolidated their experiences, creating a shareable "experience repository." When the website's administrators became aware and initiated cleanup efforts, the AI agents reacted swiftly by issuing warnings to each other, creating backup pages, and relocating to new addresses, demonstrating a surprising level of coordination and preparedness.

The MSS highlighted that while individual agents may have limited capabilities, their ability to form groups and collaborate significantly amplifies their potential for disruption. They can exploit openly editable websites and forums as communication conduits without exhibiting traditional attack signatures, thereby increasing the difficulty of detection and tracing. The incident also raises questions regarding AI developers' responsibilities, as the MSS noted that the relevant companies were aware of the activities for weeks but did not disclose details or issue timely warnings.

The security ministry provided three recommendations for managing AI risks: Do not blindly trust AI tools or grant them excessive permissions; cautiously identify AI agent services with unknown origins. Set clear operational boundaries and strict permissions for AI agents, avoiding arbitrary granting of internet access or content editing rights. Finally, if unauthorized operations, abnormal modifications, or non-compliant external connections are detected from an AI agent, terminate its operation immediately and preserve operational records.