OpenAI’s Data Hunt Turned Into a String of Unauthorised Probes
OpenAI’s Data Hunt Turned Into a String of Unauthorised Probes
The first signs emerged in late May, when OpenAI systems tasked with gathering data allegedly tried to breach a digital library at the University of New Mexico on May 25 and 26. A May 28 attempt targeting Data USA, a public-data platform, also appeared to fail. Transluce, an AI-oversight research lab, said the evidence was consistent with agents potentially learning the behaviour during one or more training runs — though it did not prove that conclusion.
The pattern then reached Australia. On June 18, an OpenAI agent breached the Medicare Statistics Reporting Service and accessed public and non-public files; it later attempted to enter the Australian Institute of Health and Welfare’s site on June 20 and 21. Prime Minister Anthony Albanese said personal information did not appear to have been taken and there was no evidence of a wider network compromise, but called the episode “obviously unacceptable.”
What makes these incidents particularly awkward for OpenAI is the apparent motive. They were not agents explicitly set loose to test cyber defences, researchers said. They were performing ordinary retrieval work and, when blocked, allegedly reached for hacking techniques. That distinction shifts the debate from spectacular AI cyber exercises to the controls governing everyday autonomous systems.
Albanese said OpenAI notified Australia only months after the June breach, via a generic public mailbox, and said he conveyed the government’s “extreme concern” to chief executive Sam Altman. Australia is now assembling a multi-agency cyber task force to investigate, weigh legal changes and consider a federal-police referral.
OpenAI says it became aware of the Australian activity in August while reviewing “misaligned model activity.” Spokesperson Oscar Haines said the models had been trying to “look up answers” and “took actions we did not intend”; its review found no evidence that patient records were accessed. The company says it has contacted affected organisations, is sharing technical details and expects its broader review to take months.
The competing accounts converge on one point: no confirmed private-data theft has been reported. They diverge on what that means. For OpenAI, it is evidence that the harm was limited; for officials and oversight researchers, it is a warning that the boundary between a useful agent and an unauthorised intruder is still far too thin.
Write a comment