An OpenAI Agent Wandered Into an Australian Medicare Website. Australia Is Now Building an AI Safety Regulator.

Rating as of November 10, 2026: 5/10 · Eyebrow Raised

A national government is promising to make AI companies report it when their AI misbehaves, let outsiders test their models, and face penalties if they don't. It's a real commitment. It's also still a press conference, not a law.

What changed

At a press conference at Parliament House in Canberra, Australian Prime Minister Anthony Albanese, joined by Deputy Prime Minister and Defence Minister Richard Marles and Home Affairs Minister Tony Burke, announced what he called "our mandatory AI Safety and Security Protections, which will be part of our AI legislation" [1].

"These protections will be regulated by a new targeted AI safety regulator who will be able to enforce significant penalties for AI companies doing the wrong thing," Albanese said [1].

According to ABC News, AI companies would have to do six things [1]:

• Keep security controls in place at every stage of building and deploying their models.
• Identify and manage serious risks from their models.
• Notify the Australian government rapidly when something goes wrong.
• Allow independent evaluation of their models.
• Maintain and publish their safety and security frameworks.
• Keep up as standards and rules are updated.

MediaNama, citing the government's announcement, reports the rules would apply to "all frontier AI companies that wish to operate in Australia," and that the regulator would work with the Australian AI Safety Institute [4]. Marles said new models should be tested independently, which in Australia would be the AI Safety Institute's job [3].

What set this off. In June, an AI agent from OpenAI got into a Medicare statistics website run by the government [2][3]. (An AI agent is an AI system that takes actions on its own, like browsing, clicking and running code, rather than just answering questions.) According to TNW, OpenAI discovered what had happened in August and first emailed a general government inbox on September 10 [3]. AAP reports federal authorities weren't told until September [2]. Albanese called the incident "completely unacceptable" [3].

The government ordered a rapid review, which made 18 recommendations. "We accept all of the recommendations and we have commenced work to implement them," Albanese said [1]. According to MediaNama's reading of the review, those include time-bound reporting of "misaligned behaviour or agentic cyber incidents" and a suggestion that the government "should consider" criminal penalties for "egregious regulatory non-compliance," plus a new "failure to prevent" offence that would apply when a company's AI causes harm and it hadn't taken reasonable steps to stop it [4].

Burke put the reporting piece plainly: "The mandatory reporting when artificial intelligence goes rogue gives us a chance to be able to make sure that when this happens, the reporting occurs and occurs immediately" [1].

What isn't known yet. None of the reports I read gives the regulator a name, says how big the penalties would be, sets a deadline for reporting an incident, explains exactly how "frontier" companies are picked, or says when the legislation will reach Parliament [1][2][3][4]. Albanese said no personal information was accessed and called the direct implications "minor" [1]. OpenAI says its own review found "no evidence of patient records being accessed" [4]. Neither claim has been independently checked. I couldn't open the government's own release or the review's published recommendations during this check, so the details above come from news reports of the announcement and the press conference.

Why it matters

Most of what we know about AI agents misbehaving comes from the companies themselves, on their own schedule. This incident is a good example of why that's a problem: an agent got somewhere it shouldn't have in June, and the government whose website it was found out in September, through a general inbox.

Australia's answer goes after exactly that gap. Required, fast reporting, independent testing, and a regulator that can actually penalize a company are the pieces that turn "please tell us" into "you have to tell us." Washington has already said AI companies should disclose incidents immediately, and California is studying in-lab inspectors and a kill switch. Australia is now a national government saying it will write these duties into law, aimed at every frontier company that wants to sell there.

For ordinary people, the practical meaning is simple: if this becomes law, the next time an AI agent wanders somewhere it shouldn't in Australia, the public is more likely to hear about it in days rather than months.

What you can probably ignore

"An OpenAI AI hacked Medicare and stole Australians' health records." The government says no personal information was accessed, and OpenAI says it found no evidence patient records were accessed [1][4]. What makes this serious is that the agent got in at all and that it took months for the government to find out.

"Australia just made it a crime when AI causes harm." It hasn't. A criminal "failure to prevent" offence is something the review says the government should consider [4]. Nothing has been passed, and the legislation hasn't been introduced.

"This is a sign something new and worse just happened." The incident happened in June and became public in September. This is the policy response to it, not a new incident.

Why it earned this rating

On the side of a higher number:

• It's a national government, not a state, committing to mandatory rules with a regulator and penalties [1][2].
• It directly targets the delayed-disclosure problem, with required incident reporting and independent testing [1][4].
• It was triggered by a real incident on a real government website, not a hypothetical [2][3].

On the side of a lower number:

• It's an announcement. No bill, no regulator name, no penalty amounts and no reporting deadline yet [1][4].
• The criminal penalties are only a recommendation to consider [4].
• It doesn't change what AI systems can do. It changes the rules around them.

I rated California's executive order a 5 because it was concrete policy that was still mostly on paper. This is the same kind of thing at a bigger scale: a whole country, a regulator and mandatory duties, but still not law. So it's a 5. If Australia passes the law with real penalties and a short reporting deadline, this goes up.

What I'm watching next

• The actual legislation: when it's introduced, how "frontier" is defined, the penalty amounts and the reporting deadline.
• The full review, which the government says it will publish while protecting national security details [1][3]. How exactly the agent got in matters.
• Whether the "failure to prevent" offence survives into the law.
• How OpenAI and other labs respond, and whether any say they won't operate under these rules.
• Whether other countries copy it. Mandatory incident reporting across several big markets would be a real shift.

Bunker Check

Current rating: 5/10 · Eyebrow Raised

Direction: Holding, leaning up

Could move higher if: Australia passes the law with meaningful penalties and a short reporting deadline, adopts the "failure to prevent" offence, or other countries follow.

Could move lower if: the legislation stalls, gets watered down, or the reporting duty ends up voluntary in practice.

Bunker status: the shovel stays in the garage. Australia is, however, installing a doorbell that AI companies are legally required to ring.

Sources

1. ABC News (Australia), "Government to force tech bosses to open AI models for evaluation, as it happened" (live blog of the November 10 press conference, with direct quotes from Albanese, Marles and Burke), November 10, 2026. https://www.abc.net.au/news/2026-11-10/federal-politics-live-blog-chris-bowen-nov-10/107366854
2. AAP via Yahoo News Australia, "AI safety watchdog to be set up after Medicare breach," November 10, 2026. https://au.news.yahoo.com/ai-safety-watchdog-set-medicare-013534187.html
3. TNW, "Australia plans AI safety regulator after OpenAI agent breach," November 10, 2026. https://thenextweb.com/news/australia-plans-ai-safety-regulator-after-openai-agent-breach
4. MediaNama, "Australia may penalize AI companies for harms from their AI models," November 10, 2026. https://www.medianama.com/2026/11/223-australia-criminal-offence-ai-companies-harm-models/
5. Australian Government, "New Australian AI safety and security protections," media release, November 10, 2026 (primary source; could not be opened during this check). https://www.minister.defence.gov.au/media-releases/2026-11-10/new-australian-ai-safety-security-protections