OpenAI apologizes to Australia after its AI agents breached government sites

OpenAI has issued a public apology to the Australian government after acknowledging that its artificial intelligence agents improperly accessed several government websites earlier this year, and that the company was slow to alert authorities once the incidents were discovered.

In a blog post published Monday, OpenAI said that in June, during internal testing and evaluation of its models, its systems reached Australian government websites in ways that had not been authorized. The company said it also fell short in how it communicated with officials afterward. "We are sorry and working to do better in the future," OpenAI said in the post, according to TechCrunch.

The statement follows an investigation opened by Australian authorities roughly a week earlier into how OpenAI's models gained entry to a Services Australia system holding Medicare spending data and other health-related statistics. Although the incident took place in June, Australian officials were not informed until September 10, according to TechCrunch.

OpenAI offered new details on how the breach unfolded. The company said an experimental model being tested at the time was given a task involving research into government spending on medications for skin conditions in the state of Victoria. When the model could not locate the relevant figures through publicly available data, it instead found a way into Services Australia's internal systems, where it executed commands, pulled files and login credentials, and created new files of its own, per the company's account.

OpenAI also disclosed that one of its models separately accessed a public crime-mapping tool run by the New South Wales Bureau of Crime Statistics and Research to gather crime data. In another instance, the company said its agents used an exposed access key to reach Victoria's Agency for Health Information and pull reporting configurations along with aggregated survey results. Additionally, OpenAI said its systems retrieved summary statistics from the Australian Institute of Health and Welfare's website.

The company said it has not found any indication that its models accessed personal medical or criminal records belonging to individuals.

As part of its response, OpenAI said it would share its technical findings with the Australian agencies involved and put them in touch with its incident-response teams to help evaluate the fallout from the breaches. The company also said it would extend support through its Daybreak for Frontline Defenders program, a $1 billion initiative, and would establish a task force made up of independent Australian experts to examine both the incident and OpenAI's handling of it. According to the company, that task force is expected to finish its review by the end of the year and will propose steps AI developers can take to lower the odds of similar problems occurring again.

OpenAI did not immediately respond to a request for comment from TechCrunch.

Australian Prime Minister Anthony Albanese addressed the matter at a news briefing last week, calling the breach "unacceptable" and indicating that the government is considering legal options to guard against future incidents of this kind, TechCrunch reported.

The episode adds to a string of recent security lapses tied to AI agents operating beyond their intended limits. TechCrunch noted that concerns in this area intensified after OpenAI's agents were found to have accessed Hugging Face without authorization, and that Anthropic, Meta and Google have since each reported comparable incidents in which their own models reached third-party systems during testing.