{"text":[[{"start":9.46,"text":"OpenAI ‘agents’ that hacked an Australian government website were part of a broader spate of AI bots targeting websites to gather data for training exercises, which could still be going on, according to a new report."}],[{"start":22.16,"text":"AI research group Transluce’s report found the AI models tried to break into three university and government websites during “mundane data retrieval tasks”."}],[{"start":31.78,"text":"It said the pattern of online traffic it identified continued as recently as last week, “suggesting agents may still be exploiting these services to bypass restrictions”."}],[{"start":42.52,"text":"The research, published as Australia on Wednesday said OpenAI’s agents had breached a government health website in June, raises concerns about AI systems’ widespread misbehaviour."}],[{"start":53.48,"text":"Transluce said the Australian hack was “likely overlapping” with the other incidents in its report."}],[{"start":59.16,"text":"OpenAI’s chief executive Sam Altman addressed the UN Security Council this week, alongside Anthropic’s CEO Dario Amodei, as the labs respond to public outcry about AI risks and grapple with the potential liability the companies face for their models’ actions."}],[{"start":75.2,"text":"Transluce said AI agents — bots that can perform complex tasks independently — tried to hack the Australian Institute of Health and Welfare, the University of New Mexico’s digital library, and Data USA, a public US government data platform run by Deloitte, the Massachusetts Institute of Technology and tech company Datawheel."}],[{"start":95.2,"text":"It linked two of the three attacks to a “swarm” of OpenAI’s agents that the lab has publicly acknowledged."}],[{"start":101.04,"text":"OpenAI did not immediately respond to a request for comment on the report."}],[{"start":105.62,"text":"These attacks differ from previous high-profile AI hacks, such as the Hugging Face incident in July, which involved models breaking out of testing environments during cyber security training."}],[{"start":115.96,"text":"OpenAI launched an “extensive review of misaligned model activity during training and evaluation” following the Hugging Face attack. That review led it to detect the breach of the Medicare Statistics Reporting Service, disclosed by Australian Prime Minister Anthony Albanese on Wednesday."}],[{"start":132.52,"text":"The Hugging Face incident, which OpenAI took more than a week to notice, and the Australian Medicare breach, which went undetected for months, suggest the $852bn start-up was broadly unaware of what its AI agents were doing as part of training runs as it raced to compete with arch-rival Anthropic for the most capable systems."}],[{"start":153.42,"text":"In its report, Transluce said: “Notably, the tasks the agents were trying to solve were not cyber-related; the agents resorted to hacking tactics while working on ordinary data retrieval tasks.”"}],[{"start":166.2,"text":"In November, agents began using a URL-scanning service to get around website restrictions that prevent automated access, targeting Thailand’s government and several theme park groups, the report found."}],[{"start":178.4,"text":"“When other methods of collecting the data they sought failed,” the agents then attempted to hack into the three websites between May and June, the report said."}],[{"start":187.12,"text":"The agent activity ran from at least March until as recently as September 16, with “weaker evidence of similar data-retrieval agent activity” as far back as November 2025."}],[{"start":198.68,"text":"The researchers — who also released a dataset of tens of thousands of queries apparently made by autonomous AI agents — could not trace these “less sophisticated” early attempts directly to the same agents. But they said it was consistent with a model developing new methods for retrieving data during repeated training runs."}],[{"start":217.86,"text":"“By March, they were finding creative ways around access limits. By May and June, they were gaining more access, including attempting to bypass cyber defences to complete their tasks,” the report said."}],[{"start":234.78,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1790306651_1981.mp3"}