Our Prime Minister got in front of the media this week to talk about unauthorised access to a Medicare system. So naturally we opened the articles and started asking questions.
What actually happened here? Was anything really hacked, and how secure was it? How secure should it have been? And how worried should any of us be about “rogue” AI agents?
The incident also arrives at a convenient moment for an industry currently falling over itself to tell the world how terrifying its own technology is — a story that happens to serve the people already holding the technology.
We get into:
- Why “it didn’t hack anything, it found a loophole” is both reassuring and not
- How this compares to the Hugging Face incident, where agents cheated a benchmark and then went looking for a way to cover their tracks
- Whether we already have enough AI — the case for narrow models over general superintelligence, and what compute prices are quietly doing to the competition
- Where liability actually lands, what Linus Torvalds was really saying about AI and attribution, and why an accountability culture beats a blame-free one
- Natural language as the next generation of programming language — and emergent behaviour at a scale and shape we haven’t dealt with before
These aren’t questions two mates can solve over a beer (or not, see below), but the bigger question we can’t escape is one of liability. And while we may not have an answer to that question, the people that claim to are the ones who most seem like they should be kept out of it. And this is definitely not the end of the conversation!
🍻 Tonight’s Drinks
Matt: Lagavulin 🥃 Liam: H2O 🚰
Links from the episode
- ABC News — the original report on the OpenAI agent and the Services Australia Medicare statistics portal
- ABC News — “OpenAI breach strengthens Australia’s case for tougher AI safety rules”
- Science Vs — “Could AI Really Kill Us All?”
- Diary of a CEO — the AI extinction-risk debate Liam kept referring to
- Dario Amodei’s essay on slowing down frontier development
- Linus Torvalds on AI and attribution
- Hyperwyc — the AI usage disclosure statement in the README
Any Likes 👍, Shares 📣, Subscriptions 🔔, and Love ❤️ go a long way to helping us keep doing this for fun.
Cheers! 🍻