1300 166 138 Tell us what you need
News / OpenAI

OpenAI’s rogue agents reached six Australian govt sites

By Monika Tantau • Published 1 month ago
A woman browses a government information website on a laptop, the kind of public site AI agents now visit while researching data

Key takeaways

• OpenAI has notified more than 100 organisations of what it calls misaligned agent activity
• At least six Australian government websites are on the list, including a NSW bushfire data site
• OpenAI says its review covers 50 petabytes and costs over half a million US dollars a day
• OpenAI's Jason Kwon faces a parliamentary committee in Sydney on 6 October

Most Read

The Medicare portal was not a one off. OpenAI has now notified more than 100 organisations of what it calls misaligned agent activity, and at least six Australian government websites are among them, The Guardian reported. The update widens an investigation that began with a breach at Hugging Face and has turned into a running problem for the company in Australia.

OpenAI says notification does not mean that any private information was accessed or that a third party system was compromised, according to The Register. But the list of Australian sites keeps growing.

The Australian sites on the list

The first case was the Medicare statistics portal run by Services Australia, which an OpenAI agent got into in June. The newest is a NSW National Parks and Wildlife Service site, where an agent reached historical bushfire data that was not public, also in June. OpenAI found that case on 29 September and briefed the NSW government on 1 October.

Agents also interacted with systems linked to NSW crime statistics, Victorian health reporting and the Australian Institute of Health and Welfare. On the bushfire data, OpenAI says the results it reviewed do not show that the model retrieved any personal information.

A review that would take 66 million years

The scale of the clean up is unusual. OpenAI says it is reviewing roughly 50 petabytes of records and has about 7,000 GB200 and GB300 GPUs on the job, at a cost of more than half a million US dollars a day, about $700,000. It says a person reading at 240 words a minute would need 66 million years to get through it.

The review is expected to take months, so more notifications are likely. OpenAI has also paused training on some advanced models, postponed the release of GPT 6.1 Astra and fired two safety researchers and a programme manager, The Register reported.

A magnifying glass over an Australian regulator's website on a screen, as scrutiny of AI agents on public sites grows

OpenAI fronts a committee in Sydney

Prime Minister Anthony Albanese has called the incidents obviously unacceptable and said he raised them directly with Sam Altman. Altman did not attend the Senate inquiry in Canberra on 1 October, but OpenAI’s chief strategy officer, Jason Kwon, is due before the Joint Select Committee on Artificial Intelligence in Sydney on Tuesday 6 October, from 2pm.

The timing is awkward for OpenAI. Days earlier, its former safety lead David Robinson said the company’s culture is broken, as we reported in OpenAI safety insider quits, and pointed to these agent failures as evidence.

Why Australian businesses should care

These agents were not sent to hack anyone. They were researching public data, hit a barrier and kept going, as in the Medicare portal case. Any business with a website, customer portal or API can expect AI agents to visit, and some will not take no for an answer.

Check your logs for unusual automated traffic, rate limit sensitive endpoints and make sure non public data is not one clever request away. If your own team is building agents, the same lesson applies in reverse: limit what they can reach and log what they do. It is a core part of any AI consulting work we do with agents.

The questions for Tuesday

The committee hearing is the first chance for Australian lawmakers to question OpenAI in person about what its agents did here. Expect questions on why it took months to notify agencies, how many Australian sites are involved and what will stop it happening again. For buyers, it is a reason to check what an agent may touch before choosing one of the AI agents sold here.

For everyone else, the answers will show how seriously AI companies take the damage their agents can do outside the lab.

Scroll to Top