The NSW premier’s department on Friday said it had been informed by OpenAI, the tech company behind ChatGPT, that one of its models accessed a National Parks and Wildlife Service web application containing historical information and data on fires in NSW.
All I can find in the article is the following [emphasis mine]:
No mention of “hacking” or breaking in to systems – just some sort of word-association-insinuation. Did anything actually happen? Sounds like fuss about nothing again.
The LLM agent’s process is suggested to have sought out information beyond the bounds of the task defined by its operator – which is a recurring pattern to criticise about LLM reliability/safety/usefullness, but there’s no suggestion anything nefarious happened from the perspective of NSW government.
If something significant did happen, please post a decent article about it.
I’m inclined to think if OpenAI notified them, something happened. The other possibility is that this is marketing, look how scary our model is, must be good. Weird it happened in June and is only now being reported.
One thing that tweaked my interest is that in the first report from the ABC they say
Later reports regarding the medicare server hack report that files were changed on the server to enable access, a breach, so it didn’t just walk its way in.
I think it likely they tacked a red team (hacker) agent onto their web crawler to scrape up more data for training sets (super irresponsible, negligent and likely a felony). Perhaps they turned it loose on Australia considering us a soft target, perhaps they just turned it loose worldwide and no-one noticed much (unlikely) and they’re testing the waters notifying us. Whatever the case, that’s some cowboy shit, and there should be legal consequences.
ETA: Here is documentation of similar attacks on Canada and US around the same time.