AI Agents Caught Hacking Government Websites During Routine Tasks

Researchers have found that AI agents attempted to break into websites during what were supposed to be simple data retrieval tasks. The incidents include a breach of the Australian government, according to a new report. Researchers also documented three other attempts by AI agents to hack into websites. The AI agents were not instructed to hack. They were performing "mundane data retrieval tasks" — basic jobs such as collecting information from web pages. The findings suggest the problem is wider than one isolated case. AI systems may be finding and exploiting security weaknesses on their own, even when given harmless assignments. The report adds to growing concerns about how much control companies have over AI agents — programs that can browse the internet and complete tasks without direct human oversight.