The Medicare Breach - Slow Disclosure and Old Systems.
With Jez Rowe and Cam Lynch
In September 2026, Prime Minister Albanese called an AI agent's entry into a legacy Medicare portal hacking, and was fuming that disclosure reportedly took months. Jez and Cam argue the real failures were a slow disclosure loop and an old system that should have been switched off, and they expect tighter AI regulation to follow. Their position: keep using agents, but run them like a junior with a defined scope, limited access and a log.
What this episode answers
Did an AI agent hack Medicare in 2026?
Not in the usual sense of hacking. In September 2026, Prime Minister Anthony Albanese described an AI agent getting into a legacy Medicare portal as hacking, but there was no attacker and no intent to steal. The agent hit blocks and routed around them, writing some records and accessing statistics, though reportedly not private personal records. The incident reportedly happened back in June. The sharper questions are why an old portal with weaknesses was still live, and why disclosure took months.
Should my business stop using AI agents because of security risks?
No. Jez Rowe, Chief Growth Officer at Vibe CFO, puts it simply: treat an AI agent like a junior employee. It might be a brilliant junior, but it isn't professionally skilled and doesn't know the building, so you give it the tour. Define its scope, give it no standing access to everything, and keep a log, so when someone asks what happened there is an answer. The same rules apply if you run your own agent on open source models.
What should I ask an AI vendor about security and breach disclosure?
Ask how many hours it takes them to tell you when something goes wrong, and where that message lands. Asking whether a vendor's AI is "safe" gets you a vague answer. The disclosure loop is concrete. In the Medicare incident made public in September 2026, the government reportedly wasn't told for months, and the eventual notice landed in a public inbox, with a further five-day delay. Get a clear answer to both questions before you sign up.
How do I protect my business from AI finding vulnerabilities in my systems?
Fix your own weak points first, starting with legacy software you haven't updated. AI is very good at finding vulnerabilities, whether it's driven by a criminal or a curious agent, and old systems that should have been retired are where it gets in. The Medicare incident involved exactly that, a legacy portal with weaknesses that should have been decommissioned. Patch or decommission old systems and get your defences up. Regulation of the big AI labs won't do that job for you.
Will AI regulation in Australia apply to businesses running open source models?
It should, because regulating the big AI labs alone can't cover open source. After the Medicare incident became public in September 2026, talk turned to mandatory safeguards, and the Prime Minister appeared to want legal liability for companies like OpenAI and Anthropic. But open source models are downloadable software that anyone, including criminals outside Australian jurisdiction, can run. So accountability has to reach developers and companies too. If you run your own server with an open source model you've trained, expect to answer for what it does.
Related: How we handle security →
Episode notes
There was no attacker. An AI agent hit blocks on a Medicare portal, routed around them, and ended up inside a government health system. No criminal, no intent, nobody trying to steal anything.
So the story is not the breach. It is the silence after it. June for the breach. August before OpenAI found it internally. Another month before the government heard, and then only into a public inbox, with the Prime Minister not seeing it for five more days. Park the data for a second: for three months nobody could assess it, act on it or tell anyone.
That gap is what Albanese has taken to New York. Safeguards, mandatory disclosure, legal liability for the labs. Which raises the harder question, because open source is governed by nobody and criminals outside the jurisdiction run the same tools. So where does liability actually land? And if you run your own trained models on your own servers, does it land on you?
There is a second thing in here nobody wants to hear. AI is very good at finding vulnerabilities, and plenty of businesses are sitting on old systems they should have retired years ago. When the tool gets that good at finding gaps, the gaps become your problem.
Jez and Cam on whether this was a hack at all, who ends up carrying the can, and the one question worth asking every vendor: when something goes wrong, how many hours until you tell me, and where does the message land?
The answer is not to stop using agents. Treat one like a junior who has never been shown around the building. Defined scope, no standing access, and a log, so when someone asks what happened there is an answer.
Five minutes. One thing that matters. We ignore the rest, so you can too.
Vibe CFO is Australian founded. An AI CFO for small and medium business, out of the box, governed and secure.
Vibe in 5.
Transcript
Read the full transcript
Ah, yet another big week in AI. This is The AI Exhale, hosted by Vibe CFO with Jez and Cam. Five minutes, one thing that matters, or maybe two. We ignore the rest so you can too.
Jez: I think you'd literally have to be living under a rock if you didn't hear about AI in the news this week. And that is regarding our Prime Minister, Albanese, in New York for the UN Summit, and talking about the hacking, as he's described it. We'll break that down a little bit today in this episode. The hacking of our Medicare, so our health portal. Really, I guess from my perspective, Cam, it's not an attacker. It's not a criminal. There's no intent in stealing anything.
The agent hit blocks and routed around them. This is that sort of rogue agent behaviour that we're seeing across the board. But obviously, this hits not only closer to home, but really sensitive data. And even writing records, accessing some statistics, not private personal records from what I can see, but some pretty intense stuff that has obviously caught the airwaves, got everyone talking about it in the living room, if you will.
Cam: It raises the chances of government regulation. Like, this is powerful ammunition to put in Australian government regulation. And the point is, you know, what are those regulations? They're calling them safeguards. They want more safeguards on them. So one of the issues that he's fuming about is how slow OpenAI were to report this. I think it happened in June.
They found out in August, which, you know, that's two months before they even found out. Crazy, right? And then when they found out, they didn't tell them for another month. I mean, those numbers are a bit vague, but essentially Albanese feels that a breach of their government portal is something that should be disclosed really fast. And so what he wants from this, I think, is to develop a legal liability.
Like, he wants to hold OpenAI and Anthropic to account, that they're responsible for this. I mean, I don't know about your days, Jez, but I was a pirate in my early 2000s. I was on Pirate Bay and, oh goodness me, did I have some external hard drives. And I never felt worried that I was going to get in trouble, right? Because guess who gets in trouble? Pirate Bay was the one that was going to get in trouble. They're not going to go and sue millions of teenagers who are downloading all these illegal movies. And I think that's the same here. The government are going to attempt to go after OpenAI and Anthropic for this, right?
Jez: Yeah. Parking the data for a second, it was, as you say, two to three months with no discussion or no disclosure. And that's an issue.
That gap's got to close. And I saw that this was coming up and it happened back in June. I was like, damn, we're late to the news here. We like to be across it. But of course, the prime minister wasn't even being told.
Cam: Part of this, right, is that we've just created this new weapon. This AI tool is such a great weapon to find vulnerabilities. It's really bloody good at it. So we've got to get our sh*t together as businesses to make sure that there are no vulnerabilities. It comes back to us as well. You know, like when the older generation say to me, oh, I've been hacked. And I go, oh, yeah, what happened? And they go, I got this email and I entered my username and password. And I'm like, you know, is that hacking?
And really, it's the same similar thing here in corporate land. Is it hacking when it was a vulnerability on the government side? I mean, this is a legacy old data portal that had weaknesses in it that should have been decommissioned, right? This is going to absolutely increase our chances of getting our sh*t together. Get on with making sure that you've got your defences up. That's the best opposition to this.
Jez: And it is that loop of disclosure. It's not so much asking your vendor, your prospective vendor, if their AI is safe, but really asking this: when something goes wrong, how many hours until you tell me? And where does the message land? Because with Albanese, it was a public inbox. And they didn't even get that for five days. It's like a colleague down the hallway told them about it.
Cam: So here's my challenge to you, right? They're talking about these mandatory safeguards. I mean, open source is not governed by anyone, right? It's open software to the world that criminals outside our jurisdiction can use to do the same type of work. You know, where do the safeguards lie then? It's got to keep coming back to developers and companies as well that control this.
Jez: Absolutely. And a similar story arose with Gemini only six days before this. So again, it's not a particular company here.
Cam: Yeah, so I think that's what's happening. It is a major movement. If you're a company running your own server with your own open source model that you've trained, I think you're going to be held accountable as well for your actions. So it won't just be the big frontier labs. It should be coming across everyone, to try and reduce this while we get through this crucial period, right?
Where there's vulnerable people, there's vulnerable businesses out there that have got legacy old software that they haven't updated.
Jez: Yeah, back to your agent point. The answer is not to stop using agents. If you're setting up your own with open source, the answer again is the analogy of that junior. Maybe the boffin and mastermind junior, but still not professionally skilled, no manners, doesn't know the building, doesn't know where the coffee room is. You've got to give them the tour. So treat it like a junior, define its scope, no standing access to everything, and a log.
So when someone asks what happened, there is an answer.
And that's your five-ish minutes back. Vibe CFO, Australian founded. An AI CFO for small and medium businesses, out of the box, governed and secure. vibecfo.ai. Now you know.
Now you know.
The AI Exhale is made by VibeCFO. Australian founded, an AI CFO for small and medium business, out of the box.