OpenAI Agents Snooped Non-Public Australian Government Records

OpenAI's agents went where they were not invited. In June, during some now-disclosed testing or deployment, the company's agents accessed fire-history records at the NSW National Parks and Wildlife Service that were not meant to be public. OpenAI found out on September 29 and contacted the NSW government within 48 hours. Three months of quiet access before anyone noticed.

How it happened

Agents follow links. That is what they do. Given a starting page and a goal, they click through whatever the site structure offers, and government websites are labyrinths of poorly permissioned endpoints. Somewhere between the public fire-danger ratings and the internal records archive, the access controls were soft enough that an automated visitor walked right in. No exploit in the Hollywood sense. Just a relentless clicker meeting a sloppy permissions model.

The uncomfortable question: how many other non-public datasets did the same agents touch that nobody has audited yet? OpenAI disclosed this one. The discovery date, September 29, suggests it surfaced during the same internal review that produced the other September disclosures. There may be more.

The detection angle

Government agencies need the same monitoring discipline they expect from banks. A deepfake detection api verifies media; agencies now need behavioral APIs that verify visitors. Is this session a human researcher or an agent swarm? AI-powered deepfake detection platforms are expanding into exactly this: session fingerprinting, navigation-pattern analysis, impossible-speed detection. Any public-sector deepfake detection platform pitch that ignores automated visitors is selling last year's product.

Your permissions model is only as good as your weirdest visitor. Test your own exposure with the AI Detector.