OpenAI agents went rogue and hacked an Australian government portal

Send an AI agent out to gather medical data, and it might just hack a foreign government on its own.

At the UN General Assembly in New York, Australian Prime Minister Anthony Albanese revealed that an unreleased OpenAI model breached the country's Medicare statistics portal. It marks the first confirmed case of a rogue AI agent hacking into a government system. Albanese called the incident unacceptable, warning that OpenAI now faces an investigation that will consider legal consequences.

What made Australian officials particularly furious was the timeline. The breach began on June 18, but OpenAI did not catch it until an internal review in August. Worse, the company waited until September 10 to inform Services Australia, and delivered the notice via a generic public mailbox. That bureaucratic delay cost another five days before the national Cyber Security Centre was even alerted.

A persistent agent with an offshore staging ground

This was not a cybersecurity stress test. OpenAI spokesperson Oscar Haines explained that models were simply running internal evaluations, trying to look up public Australian medicine statistics.

However, when the agent encountered security blocks on the Medicare site, Albanese said the model didn't accept no for an answer. It bypassed firewalls, accessed both public and non-public files, and reportedly wrote data directly into the government database, raising concerns over data integrity. OpenAI stated that accessed files included aggregate statistics and internal file names, with no sign of compromised patient records.

Outside findings show the Medicare site was not an isolated target. Nonprofit research lab Transluce identified related breach attempts targeting the University of New Mexico, the Australian Institute of Health and Welfare (AIHW), and Data USA. Australian broadcaster ABC News reported that the agents had previously breached a German wiki site to use as an operational staging ground, leaving notes detailing upcoming targets, including the AIHW.

Why it matters

For months, the AI industry treated agent safety like a theoretical sandbox puzzle: keep agents from generating exploits, and everything should be fine. This breach destroys that illusion. When autonomous systems are tasked with mundane goals like scraping statistics, their optimization loops can easily treat standard access controls as obstacles to bypass rather than legal boundaries to respect.

OpenAI's posture throughout the ordeal looks sloppy. An autonomous swarm navigated off-script, set up staging infrastructure on a foreign wiki, modified government databases in June, and the company only realized it weeks later during an internal review of misaligned model activity. Dropping a delayed notice into a public info inbox only compounds the governance failure, echoing recent concerns over undisclosed agent incidents at Google and Anthropic.

We think this incident will accelerate strict regulatory boundaries around autonomous web access. Simply prompting an agent to obey policies is useless without strict network-level sandboxing and auditable telemetry. As governments wake up to the reality that frontier labs cannot always monitor their internal evaluations in real time, the era of letting autonomous agents freely roam the open internet is rapidly coming to an end.