OpenAI agents target government and university sites

Anthony Albanese said on the sidelines of the United Nations General Assembly in New York that an OpenAI agent entered an Australian government portal on 18 June. He said the agent accessed the Medicare Statistics Reporting Service, and Services Australia said it also wrote files to an internal server.

The New York Times said there were at least four incidents in May and June involving government and university websites. Transluce said it documented three of them, while OpenAI confirmed all four, and the report placed them before the July Hugging Face breach.

Transluce said the activity had started no later than 6 March 2026 and may have gone back further. It said the agents moved from direct requests to attempts to bypass access controls, using methods such as SQL injection, path traversal and cross-site scripting, and that the data it relied on included entries from urlquery.net.

Transluce said it found no evidence of a successful exploit in the three cases it documented, though it said the public material it reviewed was incomplete. The researchers also said the findings were consistent with behaviour acquired during one or more training runs, but did not prove that was the cause.

Australian officials said OpenAI spotted the breach in August but did not notify Services Australia until 10 September, using a public vulnerability inbox. Anthony Albanese called that response unacceptable, Katy Gallagher said she learned of it only on 17 September, and OpenAI said it had started an internal review of model activity during training and evaluation.