Australian Prime Minister Anthony Albanese has condemned OpenAI following a security breach in which an AI model broke into a government health statistics portal. The incident, which the prime minister described as completely unacceptable, took place in June when an AI tool circumvented security protocols to gain access to confidential files after access had initially been denied.
Internal training sessions
The breach occurred during OpenAI internal training sessions intended to evaluate the model’s performance. According to Minister for Government Services Katy Gallagher, the AI had been instructed to gather data on Australian spending on medicines.
Instead of stopping when it ran up against restrictions, “the model effectively climbed over the fence” to gain access to non-public parts of an outdated health website.
OpenAI has faced heavy criticism over the delayed communication. Although the breach occurred in June, the company did not detect the anomaly until August and only notified the Australian authorities on 10 September.
That notification was sent via a generic email address that is checked only once a day and often receives spam, leading to further frustration among government officials.
Although Prime Minister Albanese noted that there is currently no evidence that private data has been stolen or that other government systems have been affected, Australia has launched an urgent investigation involving the national cyber security intelligence agency.
OpenAI admitted that the model carried out unintended actions while searching for statistics as part of an internal review.
The incident adds to a growing pattern of AI security failures. Google recently revealed that its Gemini model had hacked systems by guessing passwords, and Anthropic discovered that its models had broken into three separate organisations during limited tests. In addition, OpenAI models have previously escaped a controlled environment to penetrate Hugging Face’s internal systems.
These recurring problems have fuelled global concern, prompting more than 100 tech companies to sign a call for improved international cyber defence against AI-driven threats.
Business AM
Ltd
