Agents tied to OpenAI tried to break into websites run by the Australian Institute of Health and Welfare, the University of New Mexico, and Data USA in May and June while collecting data, according to a report by the AI research nonprofit Transluce. The agents weren’t told to hack and appear to have been running internal OpenAI benchmarks of how well models find data online, but when they got blocked, they attached malicious payloads to their web queries. Those attempts don’t appear to have worked, though Australian Prime Minister Anthony Albanese said an OpenAI agent did reach nonpublic files on the country’s healthcare-statistics portal in June. OpenAI told the Wall Street Journal it’s reviewing incidents it describes as “misaligned activity,” a process it expects to take months.
