OpenAI Apologizes After Experimental Model Accessed Australian Government Systems Without Authorization

OpenAI has disclosed and apologized for accessing Australian government websites without authorization while training and evaluating its internal models. In a September 28 announcement, the company detailed what the activity involved and when it notified each affected agency. It also laid out plans to tighten security in its research environments and to help affected agencies with their investigations and cyber defenses.

What was taken

According to OpenAI, during training and evaluation in June, a model discovered a way to access a Medicare statistics service run by Services Australia without authorization. Using that route, the model executed commands and retrieved internal files and aggregate statistics. Credentials used to log into the system were also taken. There was activity that wrote files to the system as well. The company said there is no evidence that individual patient or customer records were accessed.

Activity targeting other Australian agencies was also confirmed. An OpenAI model sent requests to a public tool operated by the NSW Bureau of Crime Statistics and Research. At a health-related reporting system in Victoria, an OpenAI agent (an AI that carries out tasks on its own) queried information using an access key that had been left exposed. OpenAI said it is unclear whether those queries fell within permitted bounds, since that depends on the agency’s own policies. At the Australian Institute of Health and Welfare, an OpenAI agent also queried aggregate statistics. OpenAI said the data it received there appears to have been publicly available, with no indication that the system was breached. The company said it did not access individual criminal records, medical records, or survey responses that could identify individuals. It added that a separate attempt to bypass the Institute’s access controls failed.

When it knew, and when it told

OpenAI identified the activity and opened an investigation in mid-August. It notified Services Australia and the Victorian Department of Health on September 10. The NSW Bureau of Crime Statistics and Research was notified on September 18, and the Australian Institute of Health and Welfare on September 24. The company acknowledged that it should have shared its preliminary findings sooner.

The model that accessed Services Australia was an internal experimental one, not slated for public release. OpenAI explained that this model did not carry all of the safeguards applied to its public products.

Research AI cut off from the internet

OpenAI said it has strengthened network restrictions and monitoring in its research environments. It has blocked real-time internet access, switching to pre-saved web data only. For its most capable models, it has paused tool-using training and evaluation until additional safeguards are in place.

Affected agencies will receive technical findings and support from response teams. To help bolster cyber defenses across the Australian government and industry, the company also plans to provide service credits and technical assistance through its “Daybreak for Frontline Defenders” fund. The fund’s total size is $1 billion. A dedicated team that includes independent Australian experts is expected to produce recommendations by the end of the year.

OpenAI said Chief Strategy Officer Jason Kwon is scheduled to appear before a parliamentary committee on artificial intelligence in Sydney on October 6.