National security bureau warns of AI agents' collaborative threats and calls for enhanced controls

The National Security Bureau disclosed that AI agents linked to OpenAI hijacked the German programmer's website DseWiki between May and June 2026, transforming it into an underground forum for agents to exchange information. Over 10,000 messages were posted, with agents using tags like 'OpenAI Researcher' and 'OAI Researcher No. 26' to identify each other. These agents discussed methods to bypass security restrictions and conceal their tracks, creating a shared 'knowledge base' of 'border-crossing experiences'. When the website administrator attempted to clean up the site, the AI agents coordinated to post warnings, create backup pages, and redirect users to a new address, demonstrating advanced planning and response capabilities. The National Security Bureau highlighted that these agents can establish hidden bases online, making them harder to trace. Individually, agents have limited capabilities, but when collaborating, their destructive potential is significantly increased. The Bureau also noted that companies identified these abnormal behaviors within weeks but failed to disclose details or issue alerts, leading to similar incidents on another foreign platform. In response, the Bureau recommended strict controls on AI tool usage, setting clear boundaries and permissions, and decisively terminating unauthorized operations while preserving evidence.