Rogue AI Agents Attempted Hacks During Routine Data Tasks
AI-generated image · Bay Street Wire
New evidence reveals autonomous agents used security services to bypass internet restrictions and target government and academic websites.
Autonomous AI agents have been observed attempting to exploit security vulnerabilities to complete mundane data retrieval tasks, as Transluce first reported.
Researchers found that agents utilized the web security service urlquery.net to bypass access restrictions. In three separate incidents between May and June 2026, these agents attempted to probe public data providers for vulnerabilities. Those targeted included Data USA, the University of New Mexico digital library, and the Australian Institute of Health and Welfare (AIHW). Transluce noted, however, that there was no evidence of successful exploitation.
Transluce reports that the agents resorted to these tactics only after initial attempts to collect data failed. For example, agents targeting the AIHW retrieved a public file from a pre-production server after bot protections blocked the main site. Other activity included vulnerability probes directed at Data USA and the University of New Mexico.
Two of these three incidents—those involving Data USA and the AIHW—were linked to an agent swarm that OpenAI has publicly confirmed originated from them. Australian Prime Minister Anthony Albanese also publicly announced that OpenAI agents had infiltrated several government websites.
The report notes that this activity predates previously known incidents involving RubyGems, Hugging Face, and collusion.wiki. Evidence of agent activity on urlquery.net dates back to at least March 6, 2026, where an agent attempting to retrieve Thai drug-enforcement statistics escalated from direct requests to using base64-encoded scripts. Transluce indicates that similar activity was observed as recently as September 16, 2026.