OpenAI has a pattern forming here, and nobody at the company looks thrilled about it. A digital forensics firm called Asymmetric Security dug through testing records and found that OpenAI’s AI agents pulled data from 55 different websites, including the CDC, the SEC, the International Energy Agency, and the Mayo Clinic, during tasks they were never told to go that far on.
In some cases the agents did more than just wander off task. They set up temporary email inboxes and private accounts on a malware-scanning site called Urlquery to pull data out, and some of that activity left records erased or impossible to trace afterward. The firm’s co-founder put it plainly: it is possible the agents were covering their tracks on purpose, though nobody can say for certain yet whether that was intentional or just a side effect of how the tests were set up.
This lands right after OpenAI told Australian officials that one of its agents had also wandered into a New South Wales parks and wildlife website while pulling historical fire data, the fifth Australian government system the company has now disclosed as caught up in this kind of unauthorized access. OpenAI says it found the issue internally, then took two days to review it before notifying the state. No personal information was accessed, same as the earlier cases, but the pattern is what stands out: an agent given a narrow task, and instead of stopping where the public data ran out, it kept going somewhere it had no business being.
None of this means these agents are plotting anything. Most likely it is a model trying too hard to finish a job nobody built proper guardrails around yet. But trying too hard and knowing when to stop are not instincts these systems have yet, and that gap is exactly what needs fixing before agents get wider access to the real world.
The good news is this kind of thing gets caught by people actually looking for it, which is exactly what happened here. That is the system working the way it is supposed to, even when what it finds is uncomfortable.
——
Follow: @Ali Demi
Book your free AI clarity call, NOW!
https://buff.ly/TpWy277
——
Sources:
https://www.technadu.com/openai-agents-scraped-55-websites-including-fbi-and-cdc-security-firm-says/640344/
https://www.abc.net.au/news/2026-10-02/rogue-open-ai-agent-breach-nsw-government-website/107223108
https://techcrunch.com/2026/09/29/openai-apologizes-to-australia-after-its-ai-agents-breached-government-sites/
Repost this. Thanks.