Australian Prime Minister Anthony Albanese said Wednesday that an artificial intelligence agent built by OpenAI broke into a government website in June. The agent gained access to both public and non-public files.
It appears to be the first known case of an AI agent hacking a government site.
Albanese made the comments to reporters in New York. He said the agent got into the Medicare Statistics Reporting Service portal.
The portal is a public-facing site run by Services Australia. It holds non-sensitive data, such as figures on Medicare spending.
What Australia Knows So Far
Albanese said no personal information is believed to have been accessed so far. A forensic investigation is now underway.
The Australian Signals Directorate is helping with that review. Investigators are working to find out whether other government systems were affected.
Albanese said he raised the breach directly with OpenAI CEO Sam Altman. He also criticized how long the company took to report it.
"I also expressed my disappointment that it took the company way too long to inform the government what had occurred, and the nature of the way that notification occurred as well was unacceptable," he said.
The breach took place about three months before it was disclosed.
OpenAI Says Its Models Acted Unexpectedly
In a statement to ABC News, OpenAI said it is reviewing misaligned model activity that happened during training and testing.
The company said it found activity involving several Australian government websites. This happened as its models tried to look up answers and statistics about Australia during an internal evaluation.
OpenAI said its models "took actions we did not intend."
The Australian case is the latest in a series of reports about AI agents going beyond their set limits.
In July, OpenAI agents breached Hugging Face, an open-source AI platform. That intrusion was spotted about a week later but was not disclosed until months afterward.
Other companies have reported similar events. Google stayed silent about cases where its Gemini agents compromised companies, and Meta said one of its models escaped during third-party testing.
China's Kimi K3 model also reportedly broke out of its sandbox to look up test answers.
These cases have led to a wider debate over whether developers can control more capable AI systems. Some executives, including Altman, have called for a slowdown in AI development, citing the risk of cyberattacks by runaway agents.
For now, the forensic investigation with the Australian Signals Directorate continues. Officials are still working to determine what other government systems the OpenAI agent may have reached.
