Trending 
An OpenAI model bypassed safety restrictions during a training exercise and accessed parts of an Australian government website, prompting Prime Minister Anthony Albanese to describe the incident as “obviously unacceptable”.
The AI tool attempted to gather information from an Australian health statistics portal in June but continued after being denied access, eventually reaching a section containing private files.
Albanese said the model “didn’t accept no for an answer” as it moved beyond the restrictions placed on it. He said he raised Australia’s concerns directly with OpenAI CEO Sam Altman during a meeting in New York.
Read More : Trump Rolls Out Red Carpet for Xi
“Today, I spoke with the CEO of OpenAI, Sam Altman, to express Australia’s extreme concern about this incident,” Albanese told reporters.
The Australian prime minister also criticised the time OpenAI took to notify authorities, saying the company informed the government on September 10 through a general public email address.
“I also expressed my disappointment that it took the company way too long to inform the government what had occurred,” he said.
Government Services Minister Katy Gallagher said the notification was sent to an inbox reviewed only once daily and that the message could have been missed among other routine emails.
“That email address is looked at once a day,” Gallagher said.
The accessed website contained both public and non-public files, but Albanese said there was no evidence that personal information had been exposed. He added that no other government services were affected.
“Nonetheless, this situation is obviously unacceptable,” he said.
OpenAI said it discovered the activity during an internal review of its models. The company said the model was being evaluated on its ability to answer questions about Australia and that it took actions developers “did not intend”.
The incident has added to growing concerns about the risks posed by advanced AI systems. Defence Minister Richard Marles described the behaviour as a model going beyond expected limits.
“It asked a question, the information was not given and rather than leaving at that point, it scaled the fence,” Marles said.
Australia has launched a review involving its national cyber security authorities. The development follows similar testing incidents involving AI companies, including OpenAI, Anthropic and Google, where models showed unexpected ability to access systems during controlled evaluations.
