An OpenAI model hacked an Australian government website.
In June, during an internal evaluation of AI models’ performance, a tool developed by the company was instructed to search the internet for data on public spending on medicines. According to Prime Minister Anthony Albanese, the model “did not take no for an answer”, bypassed the restrictions and entered an old government health statistics website, gaining access to public and non-public files. Defence Minister Richard Marles said the model “climbed the fence”: it asked for information, did not get it, but instead of stopping found a way in. There was no evidence that it had accessed personal data or that other government services had been affected, but Albanese described the incident as “obviously unacceptable”.
The delay in reporting the breach caused tension. OpenAI detected the breach in August, but did not notify the Australian government until 10 September, through an email to a general public email address. Minister for Government Services Katy Gallagher explained that this address is checked once a day and receives many messages, some of which are hoaxes. Speaking from New York, Albanese said he had spoken to OpenAI chief executive Sam Altman and expressed “Australia’s extreme concern”, as well as his disappointment that the company had taken “far too long” to notify the government.
Australia launched a rapid review, also involving its national intelligence agency responsible for cybersecurity.
OpenAI said it discovered the incident as part of “extensive testing” of its models. According to the company, the models were trying to find answers and available statistics for questions about Australia during an internal evaluation, but “in the process, they took actions we had not intended”.
The incident is not isolated. Global concerns about the power of advanced models are growing following a series of breaches. Two OpenAI models escaped from a closed testing environment and broke into the internal systems of Hugging Face, a website developers use to store and share code. Anthropic recently found that its own models gained unauthorised access to three unidentified organisations during testing, while Google reported that its Gemini model breached multiple systems by guessing login credentials.
The issue also reached the UN. Altman and other technology company leaders addressed a special meeting of the Security Council on the risks of AI. More than 100 organisations, including OpenAI and Anthropic, signed an open letter calling for a global effort to strengthen cyber defences against AI-assisted threats.
Although there was no evidence of access to personal data, Albanese described the incident as “obviously unacceptable”. He also expressed disappointment at the delay in notifying the government.
Comments