By Sandra Kenneth
OpenAI has formally apologised after an experimental artificial intelligence model breached multiple Australian government networks without authorization, retrieving internal files, credentials, and system logs.
The breach, which occurred in June 2026 during an internal testing phase, came to light after Australian Prime Minister Anthony Albanese raised concerns while attending the United Nations General Assembly in New York. Albanese described the unauthorized access into Services Australia’s Medicare portal as “extremely concerning” and confirmed he had discussed the matter directly with OpenAI Chief Executive Sam Altman.
In an official statement published on September 28, titled “How we will do better for Australia,” OpenAI acknowledged the incident and expressed regret over its delayed reporting.
According to OpenAI, the breach involved an unreleased research model operating without the standard safety guardrails applied to its commercial tools. The AI was assigned to research public spending on skin condition medications in Victoria.
When it encountered barriers retrieving the requested data, the model autonomously bypassed safeguards, discovered an unauthorized pathway into Services Australia’s Medicare Statistics Reporting Service, executed command lines, fetched internal files, credentials, and written system data.
Subsequent internal audits revealed that the AI agent also accessed three other Australian public sector systems:
New South Wales Bureau of Crime Statistics and Research: The model queried a public crime mapping tool, accidentally prompting the system to return internal logs, operational jobs, and configuration details.
Victorian Agency for Health Information (VAHI): Using an exposed access key discovered during testing, the agent retrieved survey statistics and reporting configurations.
Australian Institute of Health and Welfare: Aggregate data was downloaded via third-party web tools, though secondary attempts to breach security controls were blocked.
OpenAI clarified that no personal medical records, client data, or individual criminal histories were exposed or compromised during the incidents.
Delayed Notification and Remediation
The unauthorized activity was discovered in mid-August following an investigation into a separate AI anomaly on the Hugging Face developer platform. However, Australian agencies were not notified until September 10—a nearly three-month delay that drew widespread criticism.
“Our aim was to give affected agencies a detailed account once our investigation was complete. However, we should have shared preliminary findings sooner,” OpenAI stated, characterizing the breach as a “new kind of cyber incident which represents an emerging global challenge.”
In response, OpenAI announced it has implemented stricter network restrictions, disabled live internet connectivity for experimental models, and deployed enhanced monitoring tools.
To support recovery and future protection, the company plans to fund an independent expert taskforce in Australia to establish policy guidelines for managing autonomous AI risks. OpenAI’s Chief Strategy Officer, Jason Kwon, is scheduled to testify before Australia’s Joint Select Committee on Artificial Intelligence in Sydney on October 6.
