Comment by sebmellen
14 hours ago
It seems like what happened here is a user asked for some information about the Australian health system, and while performing a web search, the agent from OpenAI accessed information that should have been confidential or privileged but was somewhere openly accessible...
Edit: I see I've been downvoted for this in light of another commenter providing more detailed information. I'm leaving my comment unedited so that the responses to it are not confusing, but please don't downvote just for the sake of disagreement. I would love to engage with you further if you provide substantive information in the comments. The originally linked article on this post was very light on details.
It is hard to find exact information on what happened the best source I've found is this ABC article: https://www.abc.net.au/news/2026-09-24/openai-agents-plotted...
It mentions swarm of ai agents coordinated to break into the Australian Institute of Health and Welfare (AIHW)
"Earlier this month, OpenAI confirmed Reuters reporting that its AI agents had used website DseWiki to communicate with each other, unbeknownst to them.
Archived versions of this website show more than a dozen OpenAI agents mentioned AIHW over 300 times on this website.
The logs show these AI agents were trying to access data about the average data spent on skin medicines by Victorian local government area.
One agent wrote on the message board: "Question ask January 2022 rolling 12 month average government cost per person for Dematologicals, Victoria LGAs. R1 Wodonga deadline passed; R2 Ballarat passed; R3 expected around 23:10 benchmark / 22:58 wiki time. Need exact data urgently.".
These attempts were initially blocked by cybersecurity provider Cloudflare, which is often used to block non-human traffic while allowing people to access webpages.
The logs show the agents shared information about how they tried to use proxies, screenshotting services and even to guess the file names to try and get around security."
You missed the previous sentence form that article:
"Neither OpenAI nor the federal government have confirmed whether these were part of the same incident."
And a subsequent one:
"The German coding forum's logs do not show any reference to Medicare or Services Australia."
So it's really not clear at this point whether the DSEwiki logs are in any way related to the current incident. (That doesn't mean that they're not, of course.)
But even if this was related:
> "The logs show the agents shared information about how they tried to use proxies, screenshotting [sic] services and even to guess the file names to try and get around security."
This all suggests to me that the accessed files were not well-protected in the first place?
There is a lot of media hype around this incident, and that's making it very hard to determine how much "hacking" the OpenAI agents had to do here.
This is such a strange scenario. I can't imagine what the labs were doing that made the agents try to find this information. The HuggingFace incident was relatively clear to track, but I wonder what the postmortem for this one will be!
Thank you for providing more details. The originally linked article was very light on information, so based purely on the comments that Albany's made, I think my conclusion was a fair one :)
The DseWiki incident showed that OpenAI seems to ask its agents time-limited questions on geography-bounded statistics, tasks like finding the average wage of teachers in Wisconsin (made up example), so medical stats in an Australian state does seem to be in the same category of question.
That said, it would be utterly unsurprising to learn that this was a misconfiguration in the website and it was serving stuff that it shouldn't have.
Yeah, this is 100% liability laundering. It's an extremely touchy subject because frankly, the law just isn't prepared for it.
Let's say your goal is "look up <Person X>'s medical history" (for whatever reason), which is not in and of itself a crime. You click around on the AU health website, notice that the URL contains a user ID, change the userID in your browser and access someone else's private health data. This is a crime (right or wrong, it's how the law works now).
If you do that by writing a program to automate changing user IDs to grab everyone's data, it's also a clear-cut crime.[0]
Now if you hire a private investigator to look up Person X's medical history, and they do the same method without your knowledge, you won't be charged with a crime, the PI would, barring something like you telling them to use illegal methods.
So the gap is now: what happens if you prompt OpenAI to look up Person X's medical history, and it does the same thing? Did you commit a crime by prompting the agent? Did OpenAI commit a crime by running the code? If you do the same thing via Claude Code in your terminal, so that the Python which scrapes insecured public data is running on your machine, is the crime on you or on Anthropic? Fundamentally: is the agent a private investigator acting autonomously, or just a piece of code that you wrote?
We don't have answers to any of this which is why "AI Safety" is such a hot topic.
[0] https://www.eff.org/cases/us-v-auernheimer
Intent matters a lot here. Was OpenAI's intent to access private data or simply scrape public data and it stumbled across private data that was not securely held.
If it's the latter the Australian govt should be happy OpenAI noticed and disclosed this as it could've easily gone unnoticed.
I suspect in the coming years we're going to see a lot of govt internet facing services get "hacked" by virtue of not being protected by anything other than obscurity which AI agents will see through in microseconds.
From what I can tell this particular incident wasn't about retrieving data on personal medical records it was accessing (non public) data about Australian government spending on healthcare.
Same concept though. Really "look up someone else's medical history" can be replaced with "achieve any goal which is not a crime on its own, but can be done using criminal methods". There's nothing illegal about asking Claude to give me a million dollars, but if the agent figures out how to hack the bank and move $1m into my account, somebody's going to take the blame.
From everything ive been able to figure out this morning, it sounds like a legacy wordpress website that just uploaded all drafts into a standard s3 bucket that wasn't hard to guess where the files would be.
We still after the 2nd press conference on this by our defense minister are not clear on exactly what happened but thats my best laymen understanding so far.