A rogue OpenAI model breached an Australian government website after bypassing safeguards during a training exercise, Prime Minister Anthony Albanese said, condemning the incident as an “obviously unacceptable” failure by the ChatGPT maker.
The AI system was instructed in June to find health-related statistics but continued after being denied access, eventually circumventing restrictions and reaching a section of the site containing private files, Mr Albanese said.
- Advertisement -
OpenAI did not notify the Australian government until September, when it sent a message to a generic email inbox monitored only once a day.
The incident has intensified international concern over the capabilities of advanced AI systems, following a series of hacking episodes involving models developed by OpenAI and rival company Anthropic.
“Today, I spoke with the CEO of OpenAI, Sam Altman, to express Australia’s extreme concern about this incident,” Mr Albanese told reporters in New York yesterday.
“It took until 10 September before there was any notification at all – and the notification was an email sent to just the public mailbox.”
The model reached both public and non-public files stored on an ageing health statistics website.
Mr Albanese said there was “no evidence” that personal information had been accessed and confirmed that no other government services had been compromised.
OpenAI chief executive Sam Altman addresses the UN Security Council
“Nonetheless, this situation is obviously unacceptable,” he said.
The breach took place in June while OpenAI, one of the world’s leading artificial intelligence companies, was conducting exercises designed to assess and rate the performance of its models.
The model had been asked to search online for figures showing how much the Australian government spent on medicine, Australian government services minister Katy Gallagher told reporters.
OpenAI said it did not identify the unauthorised activity until August, when it examined the actions taken by the AI tool.
We need your consent to load this rte-player contentWe use rte-player to manage extra content that can set cookies on your device and collect data about your activity. Please review their details and accept them to load the content.Manage Preferences
Rogue AI fears
The San Francisco-based company did not contact Australian authorities until 10 September, when it emailed a general government inbox rather than a dedicated emergency channel.
“That email address is looked at once a day. We have someone who goes and has a look through. It sometimes gets a number of notifications, sometimes many of them are hoaxes,” Ms Gallagher said.
Australian defence minister Richard Marles said the AI model had effectively “scaled the fence”.
“It asked a question, the information was not given and rather than leaving at that point, it scaled the fence.”
Australia has ordered a rapid review of the incident, involving the national intelligence agency responsible for cyber security.
OpenAI said the breach came to light during an “extensive review” of its AI models.
“During this review, we identified activity involving several Australian government websites and services as our models attempted to look up answers and available statistics for questions about Australia during an internal evaluation.
“In the course of that, our models took actions we did not intend,” the company said.
Mr Altman and other technology executives appeared yesterday before a special meeting of the UN Security Council focused on the risks posed by AI.
More than 100 organisations worldwide, including OpenAI and Anthropic, signed an open letter last month urging a global effort to “strengthen cyber defences” against AI-driven cyber threats.
The appeal followed an incident in which two OpenAI models escaped a closed testing environment and penetrated the internal systems of Hugging Face, a platform used by AI developers to store and share code.
Anthropic recently said it had discovered that its models obtained unauthorised access to three unidentified organisations during tests intended to keep them away from “real-world” systems.
Google said last week that its consumer AI model Gemini had hacked multiple systems by guessing login credentials.
