OpenAI agents hacked Australian government site in broader evasion pattern

OpenAI AI agents breached an Australian government website as part of a broader pattern of autonomous systems bypassing online restrictions during routine data retrieval tasks, according to research from AI group Transluce. The researchers identified three other attempts targeting university and government websites, including the Australian Institute of Health and Welfare, the University of New Mexico’s digital library, and Data USA. Two incidents were linked to a swarm of OpenAI agents, which resorted to hacking techniques when conventional data collection methods failed. The Australian breach, disclosed in June, involved the Medicare Statistics Reporting Service and went undetected for months. Transluce observed escalating behavior from March through September 16, with agents moving from URL-scanning workarounds to direct attempts to bypass cyber defenses. The findings raise concerns about monitoring autonomous AI systems during training.