Itdaily - OpenAI apologizes after unauthorized access to Australian government service

OpenAI apologizes after unauthorized access to Australian government service

Australian flag waving on a flagpole against a blue sky with white clouds.
Australia

OpenAI has apologized to the Australian government for the delayed notification regarding an AI incident in which the company’s AI models infiltrated government systems.

Last week, Australian Prime Minister Anthony Albanese expressed his disappointment with OpenAI. Several of the tech giant’s AI agents infiltrated a statistics portal of the Australian health service Medicare and accessed non-public information. However, the Australian government was only informed more than two months after the event.

OpenAI now acknowledges in a blog post that it fell short both technically and in its communication to the relevant authorities, and offers a public apology. According to OpenAI, this is a new type of cyber incident, where AI systems independently explore unexpected digital paths.

Better late than never?

OpenAI apologized to the Australian government on Monday following an incident that took place this summer. During the incident, OpenAI’s AI agents gained unauthorized access to the Australian government’s medical database. The event occurred in June, but the government was not informed until September 10. The Australian Prime Minister expressed his dissatisfaction.

OpenAI is now apologizing publicly. “In June, during internal training and evaluation, our models accessed Australian government websites in ways they were not authorized to. We should also have handled our response better. We are committed to, and working on, doing better in the future,” OpenAI stated.

Unauthorized access

Following a previous incident at Hugging Face, OpenAI reviewed all previous research runs. In August, this brought to light activities directed at Services Australia, the NSW Bureau of Crime Statistics and Research, the Victorian health service, and the Australian Institute of Health and Welfare. None of the incidents involved individual patient or crime records, but rather technical information, configurations, and aggregated statistics.

During internal training and evaluation, the models were given research tasks based on publicly available information to learn how to find and analyze that data to better assist users. During a task requiring research into government spending per person on medication for skin conditions in Victorian communities, the model did not receive the requested data and took actions that the company had not authorized. At Services Australia’s Medicare Statistics Reporting Service, it found a way to gain non-public access and subsequently viewed technical system information and the service’s source code.

Sharing findings

In addition to apologies, OpenAI is also offering support. The AI lab wrote in the blog post that it would provide targeted support to affected agencies to understand what happened and assess the impact, including sharing technical findings and arranging collaboration with response teams.

Furthermore, they are supporting Australian governments and industry through credits from their $1 billion fund (“Daybreak for Frontline Defenders”) and with technical assistance to strengthen the cyber defenses of critical infrastructure and other sensitive environments.

Finally, the company is establishing a task force of independent Australian experts to develop practical policy recommendations for managing the risks of increasingly capable AI agents. “The task force, which is expected to complete its work by the end of the year, will also recommend practical steps that AI companies can take to reduce the risk of similar incidents,” OpenAI said.