Australia tightens AI rules after an agent breached a government site
After an OpenAI agent got around the blocks on a Medicare portal, Australia accepted 18 recommendations, such as mandatory reporting of AI incidents.

The Australian government released this Tuesday (the 10th) the 18 recommendations of a review carried out after an OpenAI AI agent broke into public websites, among them the statistics portal of Medicare, the country’s public health system. In a statement from Prime Minister Anthony Albanese, the government says it accepts the recommendations and has already begun implementing them. He called the incident unacceptable and the company’s notification inadequate.
What happened
According to ABC News, the access took place on June 18, during an internal OpenAI evaluation. The agent “found a way to get around” the portal’s blocks, in Albanese’s words, and reached non-public files. OpenAI said these were aggregated statistics and internal file names, with no patient data. The company noticed the case in August and only told the government in September, through an email sent to a public help inbox.
Among the accepted recommendations are: studying mandatory, time-bound reporting of AI security incidents, including near misses; evaluating criminal penalties for a company whose model causes harm without it having taken reasonable care to prevent it; and switching off or isolating government systems that don’t need to be connected to the internet. Most of them still depend on a political decision or legislation.
What this changes for your company
1. An agent does what you asked, and sometimes a bit more. An agent with access to the internet or to your systems may try another route when it hits a barrier. List the agents in use, what each one can access and who owns each one. Everything they do needs to be logged.
2. An AI incident needs a channel and a deadline. The problem in Australia wasn’t just the access: it was almost three months between the event and the notice, through the wrong channel. Define internally what counts as an AI incident, who it’s reported to and within how many hours. And put in the vendor contract how quickly they notify you, and through which channel.
3. An old system on the internet is an open door. The affected portal was old. Take inventory of your exposed systems and switch off whatever doesn’t need to be online.
In 29 years of programming, I’ve learned that the hole is rarely in the new technology alone: it’s in the old system nobody reviews and the process nobody wrote down. With agents, both of those need to be sorted out first.
Originally published on IWA7 Insights, IWA7 Digital’s blog, which is now part of Escovando Bytes.


