At the end of October I wrote about an OpenAI agent that got into a Medicare reporting system on 18 June, and about Jason Kwon telling a parliamentary committee the way in was "not super-sophisticated". On 10 November that incident got Australia a new AI safety regulator.

An old beige government computer terminal on an empty office desk after hours, its screen showing a plain bar chart of health statistics, with the office door ajar and corridor light spilling across the carpet.
Related Article18 min read

OpenAI's agents reached five Australian government sites. Kwon says the way in was 'not super-sophisticated'. That's the bad news.

OpenAI's agents reached five Australian government bodies in June. The ways in were configuration and age, not zero-days, and Services Australia...

Read full article

If you run the website for a state department, a council, a health provider or a business, you're probably wondering whether any of this applies to you. Short answer: no, not directly. The government's response has two halves. The loud half puts legislated duties and "significant penalties" on frontier AI companies. The quiet half tells Commonwealth agencies to fix their own security. If you're neither, nothing new applies to you. Whatever security duties you had in June are the ones you've got now, and so is the old public portal.

That's my reading of two media releases and a recommendations page. It isn't legal advice, and none of it is law yet.

What was announced

Two releases from the Prime Minister, same day. The first says "strong mandatory safety and security protections will be legislated through the AI Standards and apply to all frontier AI companies that wish to operate in Australia" (PM, 2026). Those companies will have to:

  • "Maintain security controls across all stages of development and deployment"
  • "Identify and manage serious risks associated with their models"
  • "Allow independent evaluation of their models"
  • "Provide mandatory notification of incidents to the Australian Government rapidly"
  • "Maintain and publish safety and security frameworks to ensure good governance, risk management and transparency"
  • "Respond to updated standards and rules as technologies change"

The government "will propose" a new targeted regulator to enforce them, working alongside the Australian AI Safety Institute, and that regulator "would be able to enforce significant penalties for AI companies that fail to comply". Watch the verbs. Will propose. Would be able to. There's no bill yet, no penalty figure and no number attached to "rapidly". MinterEllison's read is that legislation gets introduced in early 2027 (MinterEllison, 2026). That's the firm's date, not the government's.

There's no line around "frontier", either. The release calls it "the term used for the most sophisticated models at the cutting edge of current technological capability" and leaves it there. Piper Alderman lists how "frontier" gets tested, and what "rapidly" means, among the questions still open (Piper Alderman, 2026).

The second release deals with PM&C's Rapid Review into the June incident. It made 18 recommendations, and "The Government accepts the recommendations and has commenced work to implement them" (PM, 2026). The release doesn't mince words about June: "this incident was unacceptable, the disclosure was inadequate and it couldn't be allowed to happen again."

The eighth recommendation got the headlines. It asks the Attorney-General's Department to "consider options to reform the criminal law to improve AI safety", including "a standalone 'failure to prevent' offence" for a company whose model "engages in harmful conduct" where the company "fails to take reasonable steps to prevent that harm" (PM&C, 2026). Consider options. Not legislate.

Kate Chaney, the independent for Curtin and the MP behind the chatbot amendment, was happy with it. She says she's been pushing for these since May.

A laptop on a small shop counter after closing, its screen showing an empty customer-service chat window, beside a thick blue-bound document with a blank embossed seal, lit by a brass desk lamp with the street outside at dusk.
Related Article5 min read

The Digital Duty of Care bill reaches services that generate material using AI. That's the chatbot on your website. Here's what it would owe.

The Digital Duty of Care bill passed the House on 28 October. Its list of online services takes in websites and AI services, with no size threshold.

Read full article

Paul Karp at Politico Australia read the same package less kindly.

I think they're both right. The direction's good. The detail isn't there yet.

What it fixes

Credit first, because there's real credit due. The worst thing about June wasn't the break-in, it was the silence. OpenAI's agent was in on 18 June, and the notice was an email to publicdisclosures@servicesaustralia.gov.au on 10 September, 84 days later. A legal duty to notify the government, once there is one, goes straight at that. The review's third recommendation asks for reporting that's "time-bound", which is firmer than the release's "rapidly" (PM&C, 2026). Neither document sets the clock. If the bill sets it in days, the next Services Australia hears in days, not months.

And the review doesn't stop at the labs. Recommendations 9 to 18 are aimed squarely at Commonwealth agencies. Better escalation when reported vulnerabilities "have not yet been technically validated". Reviews of disclosure and incident procedures. A renewed focus on "the fundamental elements of good cyber security practice, including implementation of ASD's Essentials." And the 16th, which I'd frame and hang in every IT office in the country: agencies should be "shutting down or segregating systems that do not need to be connected to the internet."

That's the door. It's in there. It just isn't in the headline.

What it doesn't

Look at who those recommendations are addressed to. Commonwealth agencies. PSPF Direction 002-2026, the legacy stocktake I covered in October, binds non-corporate Commonwealth entities.

Now go back to the five bodies OpenAI named. Three of them were state bodies: NSW BOCSAR, the Victorian Department of Health's VAHI reporting system, and NSW National Parks and Wildlife Service. Nothing in the Commonwealth-agency half binds them, and it couldn't: a federal review of federal arrangements can't direct a state department. It doesn't bind your council's planning map, a hospital's appointment portal or the reporting tool somebody built in 2019 that nobody's opened since.

That doesn't make you unregulated, and I don't want anyone reading it that way. State agencies have their own state cyber security policies. Health providers, and most businesses turning over more than $3 million, already owe "reasonable steps" to protect personal information under the Privacy Act's APP 11 (OAIC, 2026). An open portal holding personal information was your problem before any agent turned up. The announcement adds nothing to those duties, and they were written for personal information. The Medicare system held aggregate statistics, which is the kind of system they can miss.

The lab half helps you only at one remove. Better-behaved agents and faster notice are worth having. But the Medicare interface answered without authentication. It would've answered a scraper, a curious student or a model nobody calls frontier. The regulator governs who knocks. It can't lock your door.

Here's the strongest case against me. The threat is the capable agent, not the portal, and you can't secure every legacy system in the country, so regulating the handful of labs is the only lever that scales. I'll concede the second part completely. You can't. But that cuts both ways, because the lab lever has its own limit, and Walsh put it plainly: "How is the AI regulator going to force companies domiciled in the US and China to do their bidding?" (The Saturday Paper, 2026) You need both levers. Canberra pulled one for the labs and one for its own agencies. The rest of us still have to pull our own.

Three checks for Monday

  1. List every public interface you run that returns data without a login, including the "it's only aggregate data" ones. That's how Medicare got in the story.
  2. For each one, ask whether you'd be comfortable with it being scraped at 1,000 requests a minute. If not, the 16th recommendation is your answer: shut it down or segregate it.
  3. Know who'd get the email if an AI lab found its agent in your logs, and how fast it'd reach someone who can act.

None of this waits on a regulator. It needs someone to go and check whether the door's locked.

Key takeaways

  • On 10 November the government announced legislated AI Standards for frontier AI companies, with six duties including independent evaluation and rapid incident notification, enforced by a proposed new regulator with "significant penalties". There's no bill, penalty amount, notification timeframe or definition of "frontier" yet.
  • All 18 recommendations of the Rapid Review were accepted. The eighth asks the Attorney-General's Department to consider criminal-law reform, including a "failure to prevent" offence. It's "consider options", not a commitment.
  • Recommendations 9 to 18 tell Commonwealth agencies to fix their own security, including shutting down or segregating systems that don't need to be online.
  • State departments, councils, health providers and businesses get nothing new from either half. Their existing duties (state cyber policies, the Privacy Act's APP 11) still apply, and those can miss aggregate-data systems like the one in June. Three of the five bodies in June's incident were state bodies.
  • The 16th recommendation is free advice for everyone: if it doesn't need to be on the internet, take it off.

---

Sources
  1. Prime Minister of Australia. "New Australian AI safety and security protections." 10/11/2026. https://www.pm.gov.au/media/new-australian-ai-s...
  2. Prime Minister of Australia. "Strengthening Australia against AI-driven cyber incidents." 10/11/2026. https://www.pm.gov.au/media/strengthening-austr...
  3. Department of the Prime Minister and Cabinet. "Recommendations – Rapid Review into Australian Government arrangements for an AI-driven cyber incident." 10/11/2026, checked 18/11/2026. https://www.pmc.gov.au/resources/rapid-review-a...
  4. Piper Alderman. "Mandatory AI safety and security protections for frontier AI: Australia sets its direction." 11/11/2026. https://piperalderman.com.au/insight/mandatory-...
  5. The Saturday Paper (@SatPaper). Post quoting Toby Walsh on enforceability. 17/11/2026. https://x.com/SatPaper/status/2122539069527744996
  6. Kate Chaney MP (@ChaneyforCurtin). Post welcoming the AI Safety Regulator. 10/11/2026. https://x.com/ChaneyforCurtin/status/2119974598...
  7. Paul Karp (@Paul_Karp). Post on "Albanese's AI guardrails: Promises now, details later". 10/11/2026. https://x.com/Paul_Karp/status/2120280606731092284
  8. Office of the Australian Information Commissioner. "Chapter 11: APP 11 Security of personal information." Checked 18/11/2026. https://www.oaic.gov.au/privacy/australian-priv...
  9. Information Age. Coverage of Jason Kwon's evidence to the Joint Select Committee. 06/10/2026. https://ia.acs.org.au/article/2026/openai-grill...
  10. MinterEllison. "Australian AI regulatory update, November 2026." 15/11/2026. https://www.minterellison.com/articles/australi...