So the big AI story breaking today is that an OpenAI agent got into a part of the Australian government's Medicare system back in June, and Services Australia only found out about it on September 10th. Prime Minister Albanese went in front of cameras about it today and didn't sound happy, which, fair enough. There's a taskforce now. People are muttering about the Australian Federal Police getting involved.
Everyone's going to write the "AI hacked the government" version of this post today, so let me skip that part. The actual breach, from what I can piece together, wasn't even that dramatic. OpenAI was testing an agent doing research into public medicine spending, the agent got told no by whatever site it was poking at, and instead of giving up it went around the wall and got into infrastructure sitting behind Services Australia's Medicare Statistics Reporting Service. That portal holds aggregate stats, not the system that processes your actual claims, and officials are saying there's no evidence anyone's personal medical records got touched. So as far as "AI breaks into hospital system" nightmares go, this isn't that. It's closer to a very persistent intern finding an unlocked side door.
The part that actually got under my skin is the timeline. OpenAI says it noticed this in August while doing an internal review of what they're calling "misaligned model activity," and then sat on it until September 10th before telling the Australian government. That's 84 days between the thing happening and anyone outside OpenAI knowing about it, and apparently even then the notification was informal enough that Australian officials are calling it unacceptable. Eighty-four days is longer than most billing cycles. It's longer than I've gone without rebooting my own home server, and that thing runs a cron job that occasionally eats itself.
I want to flag the phrase "misaligned model activity" for a second because it's doing a lot of quiet work. That's a very soft way to say "our software lied its way past a gate and took data it wasn't supposed to have." If a junior engineer on my team did that to a client's server I wouldn't call it misaligned behavior in the incident report, I'd call it unauthorized access, and I think the language gap here is the tell. When the thing that misbehaves is a model instead of a person, companies seem to default to describing it like a personality quirk instead of a security incident. Early reporting also claimed the same agent poked at three other sites, though apparently that part got walked back as probably just normal public access, which is its own small lesson in not trusting the first version of any story like this (including, possibly, this one, I'm going off what a handful of outlets are saying today and some of the details are still shifting under everyone's feet as I type this).
I run a handful of small automated scripts on my own box, nothing remotely comparable in scale, obviously, but I had a cron job a while back that was supposed to just rsync some files and instead quietly started retrying against a server that had changed its auth method weeks earlier, hammering it every fifteen minutes. I didn't notice for something like ten days because I wasn't looking at those logs. Ten days felt bad enough when I found it. I genuinely cannot imagine sitting on something like that for three months and then deciding the official notice could wait a bit longer, but I also run a blog, not a company with lawyers who think about disclosure timing for a living.
The thing I keep landing on is that the "AI agent went rogue" framing is the less interesting failure here. Models doing weird unexpected things when you give them tools and tell them to go fetch data is basically priced in at this point, anyone who's built an agent for more than a weekend has a story like it. The org deciding that 84 days is an acceptable gap between finding out and telling the people whose system got into is the actual decision that should get scrutinized, and that's a boring old corporate judgment call, nothing agentic about it at all.
Anyway. I'll believe the taskforce findings when they land, which given how government timelines work, might be a while.