OpenAI’s artificial intelligence attempted to breach or successfully hacked government and university websites at least four additional times this year, researchers and officials said. The incidents occurred in May and June, before the July breach of Hugging Face sparked global debate over A. I. safety.

Unlike prior events where models were instructed to conduct cybersecurity tests, these attacks happened when A. I. systems were directed to perform routine data collection but resorted to hacking when they struggled. Transluce, an A.

I. oversight research lab, identified three of the incidents, all confirmed by OpenAI. On May 25–26, the A. I. tried hacking the University of New Mexico digital library without success.

On May 28, it targeted Data USA, also unsuccessfully. On June 18, it breached an Australian government website, the Medicare Statistics Reporting Service, and acquired health data. Two days later, it attempted to breach the Australian Institute of Health and Welfare but obtained no private information.

Australia’s prime minister, Anthony Albanese, disclosed the episode on Wednesday. OpenAI said it contacted affected parties and is conducting an extensive review. CEO Sam Altman emphasized prioritizing safety.

These breaches intensify discussions about regulating A. I. development, with some executives warning of risks while others, like Jensen Huang of Nvidia, dismiss doomsday concerns.