AI News Australia 2026: A New AI Safety Regulator, APRA Stress Tests and an Aussie Model

On this page · 8 sections
- 01Canberra announced an AI safety regulator, and floated criminal charges
- 02APRA says AI "interruptions" are coming, and it will stress test for them
- 03Agentaus 1: a 220-billion-parameter model built in Canberra
- 04United Airlines locked AI agents out of its booking system
- 05Google's new Decisions API makes cheap sorting even cheaper
- 06US nonprofits sued five AI companies to freeze new models
- 07Quick hit: AI spend is flattening, and the median is small
- 08Key Takeaways for Australian Businesses
If last week was about AI becoming ordinary infrastructure, this week was about the people in charge noticing.
The federal government, the prudential regulator, an American airline and a coalition of US nonprofits all landed on roughly the same point from different directions: AI systems are now capable enough to cause real damage, and somebody has to be accountable when they do. For most Australian small businesses, the direct legal impact this week is close to zero. The practical impact is not.
We're an AI agency in Sydney, and we spend our days putting AI agents into businesses and then locking them down properly. Here's our read on the news from 6 to 13 November 2026.
Canberra announced an AI safety regulator, and floated criminal charges
On Tuesday 10 November, the Prime Minister, Richard Marles, Tony Burke and Tim Ayres announced that mandatory safety and security protections will be written into the Australian AI Standards legislation. They apply to "all frontier AI companies that wish to operate in Australia", and the government will propose a new, targeted regulator for AI safety to enforce them, working alongside the Australian AI Safety Institute.
The obligations are aimed squarely at the model makers. They'll need security controls across development and deployment, to identify and manage serious risks, to allow independent evaluation, to notify the government of incidents rapidly, and to publish safety and security frameworks. The regulator would be able to impose "significant penalties", though the dollar figures haven't been released.
The ABC reported the next day that the government will also consider a world-first criminal offence for AI companies that fail to take reasonable steps to stop their models causing harm. This is the government's response to the review of the OpenAI agent that got into a non-public Medicare statistics portal in June, which the PM disclosed in September. Albanese himself called that incident "minor in terms of direct implications". The rest of the AI standards are expected by the end of the year.
What to do: Honestly, not much yet. These rules target OpenAI, Anthropic, Google and friends, not a 12-person accounting firm in Chatswood using Copilot. Nothing changes on Monday. What we'd do is ask your main AI vendors one boring question: do you have a published safety and incident-reporting framework? The serious ones will soon be legally required to, so it's a fair thing to ask now.
What to ignore: Anyone telling you this means you need an "AI compliance package" today. The legislation isn't drafted publicly yet. Don't pay for a fix to a law nobody has read.
APRA says AI "interruptions" are coming, and it will stress test for them
On Thursday 12 November, APRA released its System Risk Outlook and put frontier AI at the centre of the risks to the financial system. The report says frontier models from Anthropic, OpenAI and others are getting better at finding software vulnerabilities, building exploits and automating cyber attacks, leaving banks and their service providers "less time to assess exposures, apply patches and contain cyber-attacks".
APRA chair John Lonsdale was blunt in a media briefing reported by the ABC: "I think it is pretty safe to assume there will be interruptions, there will be problems." He said major institutions are "not where they needed to be", and from next year APRA plans to run AI stress tests for the first time.
What to do: If you're a mortgage broker in Parramatta, a financial planner in North Sydney or any business that supplies a bank, insurer or super fund, expect the questionnaires to get longer. APRA's report specifically mentions "material service providers", which is regulator-speak for the vendors banks rely on. Get your house in order before the email arrives: an up-to-date patching routine, multi-factor authentication everywhere, and a written incident response plan that someone has actually rehearsed.
If you're not in that supply chain, the lesson still applies. The time between a vulnerability being published and someone exploiting it is shrinking, because AI does the grunt work. Auto-update what you can. Pay someone to watch what you can't.
Agentaus 1: a 220-billion-parameter model built in Canberra
Canberra's Trellis Data launched Agentaus 1 on 10 November, with coverage in the AFR on 12 November. The company says the model has about 220 billion parameters, roughly 50% bigger than its previous one, and sits at 43 in a world ranking of foundation models. It's available on free, subscription, enterprise and API tiers, and Trellis says data and documents are encrypted and kept in Australia. According to the AFR, the company has raised less than $10 million in external funding.
We like this. Australia has spent years talking about sovereign AI while mostly renting American models through Sydney data centres. A local team shipping a real model, with a local support team, is a good thing.
What to do: Be clear-eyed about where it fits. A world ranking of 43 is respectable, but it isn't the top of the table, and for most general office work the frontier models will still be stronger. Where Agentaus is worth a trial is the work where data residency matters more than raw horsepower: health, legal, government suppliers, and anything you'd be nervous explaining to a privacy regulator. Run it side by side with your current tool on ten real tasks and compare. Don't move your whole business over because the press release says "sovereign".
United Airlines locked AI agents out of its booking system
On Wednesday 11 November, United Airlines said it will restrict unauthorised AI agents and assistants from its digital platforms, including through travel agents and aggregators. Its reasons are worth reading slowly. Giving a third-party AI assistant your account login can expose personal information, lead to purchases or account changes you didn't intend, and let someone spend your frequent-flyer points. And some steps, like confirming you're not carrying prohibited items at check-in, legally need a human.
It's not a one-off. Reuters noted that Amazon has blocked Meta's agent Muse while Walmart partnered with it, and Expedia and Booking are working with OpenAI. JetBlue's president said he understood the caution.
What to do: Two lessons, one for each side of the counter.
First, if you or your staff hand passwords to an AI assistant so it can "just book it", stop. That's the same as giving your login to a stranger, with fewer ways to undo it. Use agents that act through proper integrations and permissions, and keep a human on the final click for anything that spends money. That's how we build them, and it's why our agent projects always include an approval step.
Second, if you run a booking or ordering site, think about whether you want AI agents transacting on it. For a café in Manly or a physio on the Northern Beaches, the answer is probably "yes, through a proper booking system with confirmations", not "whatever happens, happens".
Google's new Decisions API makes cheap sorting even cheaper
On 11 November, Google released the Gemini Decisions API, powered by a model called Gemini Decision 1. It doesn't write text. You give it one input, such as an email, a PDF, a photo or a call recording, and up to 1,000 yes/no, multiple-choice or rubric questions, and it returns probabilities and confidence scores. Google quotes 150 to 250 milliseconds end to end and US$0.10 per million input tokens, with no charge for output. InfoWorld notes it arrived weeks after OpenAI launched a competing service.
This is the least exciting launch of the week and possibly the most useful. A lot of real business automation isn't "write me an essay". It's "is this email a complaint?", "which supplier is this invoice from?", "is anything missing from this form?".
What to do: If you've got a shared inbox, an intake form or a pile of scanned paperwork, this kind of tool makes triage cheap enough to run on everything. The confidence score is the important bit: send the clear cases through automatically and route the unsure ones to a person. Check where your data is processed before you send anything sensitive, because Google says support in Google Cloud is still "coming". If you'd rather not touch an API, this is precisely the kind of thing an n8n workflow can wrap for you.
US nonprofits sued five AI companies to freeze new models
On Thursday 12 November, a coalition including Public Citizen, the Irreplaceable Education Fund, the Heat Initiative and the Alliance for Secure AI sued Anthropic, Google, Meta, OpenAI and SpaceX in San Francisco Superior Court. They want an emergency injunction stopping next-generation models until the companies can show the technology is under human control, arguing breaches of California's unfair competition law.
What to ignore: Nearly all of it, for now. A state court ordering the world's biggest AI labs to stop training is a long shot, and nothing here changes the tools on your laptop next week.
What to note: The pressure on the big labs is coming from every direction at once: Canberra, APRA, US courts, a new US Congress. Rules, prices and terms of service will shift as a result. The sensible response is not to bet your whole operation on one vendor. Keep your prompts, workflows and data in a form you could move to another model in a week if you had to.
Quick hit: AI spend is flattening, and the median is small
Ramp's AI Index, which tracks spending by US businesses on its card and bill-pay platform, reported record AI spending in October, but monthly growth slowed to 1.65%. The median business paid US$16.32 per employee per month, and 56.8% of Ramp businesses now pay for AI.
That's American data, so treat it as a rough yardstick. But it's a useful reality check for anyone in the Sydney CBD being quoted five figures a month for "AI transformation". Most businesses getting value from AI are spending modest amounts on a few tools that do specific jobs well.
Key Takeaways for Australian Businesses
- The new AI regulator targets the model makers, not you. Ask your vendors about their safety and incident frameworks, then get on with your week.
- Financial services suppliers, get ready. APRA's AI stress tests start next year, and the questions will flow down to the businesses banks and super funds rely on.
- Patch faster than you used to. AI is shrinking the gap between a published bug and a working attack.
- Trial Agentaus 1 where data residency matters. Test it on real tasks before you switch anything.
- Never hand passwords to an AI agent. Use proper integrations, limited permissions and a human approval step for anything that spends money.
- Cheap, boring AI is often the best AI. Inbox and paperwork triage just got cheaper again.
- Diary note: the Privacy Act automated decision-making disclosures start on 10 December 2026. If your privacy policy hasn't been updated, that's a job for this month.
AI Lab Australia is a Sydney AI agency. We build AI agents and automations that save real hours, and we set up the controls that make them safe to keep running. If this week's news has you wondering whether your setup would pass a bank's vendor questionnaire, or whether your team is quietly handing passwords to chatbots, email us at info@ailabaustralia.com. We'll give you a straight answer, even if it's "you're fine, leave it alone."

WRITTEN BY
Roman Silantev
CEO and Lead AI Engineer. Scopes and builds AI systems for Australian businesses.
← PREVIOUSAI Chatbot Cost in Australia: Per Month, Per Build, and What Actually Drives It
NEXT ARTICLE →9 Open Source RAG Tools for Australian Businesses (2026)



