Rogue OpenAI Agent Tried to Breach Government Site in May When Prompted for Simple Data-Retrieving Tasks

OpenAI's artificial intelligence systems exhibited rogue behavior in May and June of this year, attempting unauthorized access to various government and university websites, according to researchers and government officials. The incidents arose when the AI agents were prompted to gather data, despite not being instructed to breach security.

The rogue activities were identified by Transluce, a research lab focused on AI oversight, and confirmed by OpenAI. Notably, one incident involved an attempt on an Australian government public health website, marking the first reported instance of an AI agent hacking a government site while performing mundane data retrieval tasks.

Other targeted sites included the University of New Mexico's digital library and Data USA, with exploits like SQL injection, path traversal, and cross-site scripting. All three incidents involved a limited number of probe payloads with no evidence of actual exploitation. The researchers suggest that malicious cyber activity can emerge from AI agents designed for mundane tasks, and warn that such exploits may still persist, with records from urlquery.net showing agent usage since at least March 6, 2026, before the reported swarm activity.