If you’ve been forwarded this newsletter, you can sign up free here.

Confide in me: Send your tips to this generic inbox, [email protected].

ESSENTIALS

— It’s time to sort out human responsibility for AI agents

— China’s record nuclear expansion

— Whyalla steelworks: The distressed plant has repaid more than $5 million in government grants

Listen to Playbook’s Ryan Heath on ABC’s The Party Room podcast with Fran Kelly and Clare Armstrong.

DRIVING THE DAY

HEY, WHATEVER, WE LIKE HACKED MEDICARE OR SOMETHING, LOL.

When an artificial intelligence agent goes rogue and accesses Medicare statistics, who’s responsible? The answer under Australian law is probably “no one,” Paul Karp and Finn McHugh report.

The Australian government may still refer OpenAI’s unauthorized access to police, but rather than prosecutions, the main action is likely to come from a review coordinated by the Department of Prime Minister and Cabinet closing legal loopholes.

ONE RULE FOR HUMANS: “If a person hacked into a Medicare database they would be looking at years in prison, but if you are a big U.S. tech company, you don’t even get a slap on the wrist,” Greens AI spokesperson David Shoebridge told POLITICO.

“If a physical robot entered a Medicare office and broke into a safe, that would be a crime. Why is it that [when] a digital agent does this in a digital environment we treat that differently?” asked Olivia Shen, director of the Strategic Technology Program at the United States Studies Centre.

Breathe: We cover these nuances in more detail with Nicholas Davis of the University of Technology Sydney in the Playbook interview below.

NOT AN ISOLATED INCIDENT: This is the first publicly identified instance of rogue agents autonomously breaking into government systems, but it is not an isolated incident when it comes to agents out of control.

Step back in time: OpenAI rogue agents also attempted to break into at least two American government and university websites, government officials tell The New York Times.

THE NEXT WORLD-LEADING DIGITAL POLICY? Prime Minister Anthony Albanese didn’t engineer this, but he’s been spinning around to take advantage of it, and he will mention the incident in his general assembly speech later today to reinforce his view that AI needs more guardrails.

IT’S THE AFTERMATH, STUPID: The intrusion wasn’t sophisticated, no one was harmed, but it’s still a political timebomb. The problem is what comes next — when the same thing happens but then agents start colluding to take it further, and personal data is breached or systems are shut down.

Wow! OpenAI had no idea for two months, and then had no idea how to handle this with the Australian government. Ryan Heath explains each of the major mistakes, and why this drama is part of a long history of large tech companies not taking non-American markets and governments seriously.

THE EXCUSES DON’T WASH: In Australian dollars, OpenAI has raised around $175 billion, is valued at over $1 trillion, and seeks to reshape the world. If its agents can hack the Australian health system, Playbook would be shocked if its global workforce of highly qualified and highly remunerated humans could not find the mobile number of the Australian minister’s media team. They can literally get the office phone number in an online search.

Slow: How it takes three weeks to escalate a global-first government hack to OpenAI’s senior leadership and government relations teams is anyone’s guess.

Give me just a little more time: How one meets multiple Australian ministers in that period, but there’s no system to say something, is also a mystery.

THE BUSINESS MODEL IS A CHOICE: There is no resource problem at OpenAI or its competitors. They all have the resources to ensure effective human management of their products — choosing to allocate those resources elsewhere is exactly that: a choice.

Reality check: The other reality is that San Francisco-based startups and scale-ups are not focused outside of their investors and products, let alone a smallish foreign government like Australia.

I believe in you: Australia can choose many options in reaction to those realities, the biggest mistake would be denying these realities.

WHERE DOES THIS TENSION CRYSTALLIZE? If an AI company can’t reliably know what its agents are doing, stop them when they cross a boundary and promptly tell the affected party when they discover it, what does human control over autonomous AI mean?

OpenAI CEO Sam Altman, a.k.a The One, said at the U.N. Security Council that “we will all be better off … if we can agree on what good oversight looks like.”

Hand on your heart: He may be getting the answer quicker than expected: It almost certainly involves government interference in the frictionless and low-regulation business model AI companies prize.

WHAT THEY’RE SAYING

“Damages [should be] payable by those who are most responsible,”
— Greens AI spokesperson David Shoebridge

“The government should be throwing the book at them.”
— Independent Sen. David Pocock

“Exactly the kind of canary in the coal mine we need to prompt greater action on AI guardrails, governance and disclosures of incidents.”
— Olivia Shen, director of strategic technologies at the United States Studies Centre

“Australia has few levers to protect its national interest until we are materially embedded in the frontier AI supply chain. The starting point for developing safety standards on Australia’s terms is hosting frontier AI training onshore.

“Today, Australia is relying on volunteers in the U.S. to find out about incidents where Australia is exposed. We need to urgently build accident and incident reporting capability and position to shape global safety standards.”
— Anna-Maria Arabia, CEO of the Australian Council on AI Strategy

WHAT RICHARD MARLES SAID, AND WHAT HE MEANT

What he said: “Mis-aligned activity”
What he meant: A rogue AI agent hacking government websites.

What he said: “The way we regard this is taking this incident as a salutary warning”
What he meant: Whoa, we’re lucky we got away with this one.

What he said: “Very cooperative in terms of the engagement that we are having with them now”
What he meant: OpenAI is finally playing ball after months of not taking it seriously.

BACK IN SUPERPOWER LAND — THE U.S. CHINA AI DIVIDE: U.S. and Chinese officials remain starkly divided on AI safety. Beijing is open to restrictions. The Trump administration is shooting the idea down, saying national sovereignty must trump international bodies.

WHEN YOU BECOME THE PUNCHLINE … On a night like this with Capital Brief’s ideas editor James Hennessy comes a dire personal story.

THE PLAYBOOK INTERVIEW

NICHOLAS DAVIS — “WE HAVEN’T TESTED THIS EXACT SCENARIO IN AUSTRALIAN LAW”

An industry professor of emerging technology at UTS, Davis spoke with Paul Karp on the legal issues potentially arising from yesterday’s OpenAI breach.

This interview has been edited for length and clarity.

Based on what we know so far about the incident, do you have a view about whether it breached any laws or AI guardrails?

If you conduct unauthorized access into a government system, the law says: There is criminal liability and corporations can be held to account for that as well.

So we do have laws in Australia that indicate on the surface, there would definitely be something to answer or investigate here. The trickiness and the ambiguity is that normally our laws and systems are focused on a human doing it and doing it with intent. In a situation where a company gets held liable for a human doing it, it is because of corporate culture, or acting as an agent, [the corporation] enabled it, or they were reckless in some way. We haven’t tested this exact scenario in Australian law, that I know of.

So, a genuine first?

It’s a first, that I know of, in terms of where there may not have been some kind of intent from an actor. Most of the time when we get hacks of this nature, even if they’re done by [AI] agents, they’ve been sent off to do it. Someone stands behind it saying “please go off and hack these systems.”

What level of culpability does the person who sets it off on the task have?

I don’t think we know exactly what instructions were given. I hope that’s something OpenAI can be clear about, because it gives us a much clearer idea about in the future, whether these agents are safe for organizations or you and me to use. If I’m a researcher at UTS, and I tell Claude or OpenAI to research health statistics, I don’t want to be on the line for hacking into a government server because it’s a bit aggressive.

What does this do for OpenAI’s credibility as they ask for other reforms on data centers and copyright?

I think it’s more a troubling indicator that OpenAI may not be treating this as seriously as they should. It seems strange to me that they do really care — and the Australian government cares — about getting the data center question right. We have OpenAI executives visiting Australia regularly, and an OpenAI team here in Australia as well, and yet the notification went to a generic public inbox at Services Australia. That’s an indicator they don’t have the monitoring or engagement with the government on these threats calibrated to the seriousness of the damage an autonomous agent could do.

BY THE NUMBERS

3%: The pay rise members of the Australian Defence Force will receive between Nov. 3, 2026 and Nov. 3, 2027, per a Defence Force Remuneration Tribunal decision. It’s an interim rise because the broader three-year pay rise offer for civilian defense officials is due Oct. 13.

3.5%: Consumer Price Index to July 2026.

$319 million: The estimated cost of the 3% pay rise.

$14 billion: Additional funding for Defence over the next four years, to implement the National Defence Strategy.

3 years: The length of the usual pay arrangement. (Note: Not agreement. ADF members don’t get a say).

CANBERRA MEETS WORLD

MIDDLE POWER COP: Don’t expect nail-biting, let alone successful major negotiations, at this year’s global climate summit — there’s no political appetite for it. But there is a mood among middle powers to just get on with the practicalities of energy transition, Turkey’s ambassador to Australia Ufuk Gezer tells The Australian’s Elizabeth Pike.

ONE DAY, TWO HEARTS: Fresh from signing a defense pact with Ukraine, the prime minister has inked a strategic partnership with the Netherlands after meeting counterpart Rob Jetten on the sidelines of UNGA this week.

Better than today: The agreement commits to deeper cooperation on economic and security issues — including information sharing on regional safety — and more regular meetings between leaders. The Netherlands is Indonesia’s former colonial power.

CHINA’S NUCLEAR WEAPONS EXPANSION

The Telegraph’s Sophia Yan has used satellite image analysis to identify seven sites where China appears to be expanding its nuclear weapons program.

“Taken together, experts say these images from the seven sites – Pingtong, Zitong, Lop Nur, Jiuquan, Mianyang, Jinta and Guangyuan – show that Xi Jinping has overseen a record expansion of the country’s nuclear programme,” Yan writes. This is despite comments from the leader of the Chinese Communist Party that nuclear weapons should be prohibited and destroyed.

STATE OF THE NATION

LAND AND EXPAND — CANBERRA’S CONSULTANCY OPEN SECRET

The Department of Health has awarded KPMG a contract variation valued at $499,999, more than two months into a ban on federal contracts being awarded to the troubled consultancy.

Can’t get you out of my head: It’s the latest in a series of amendments and extensions that the firm has continued to win throughout the period.

Technically, KPMG has agreed not to bid for new contracts between June 16 and Oct. 31. But extensions to existing contracts are still on the table — and the money is still trickling in. This includes a $9 million contract amendment and another worth $5.5 million, both awarded by Defence, as reported by The Canberra Times’ Zac Braxton-Smith.

I was gonna cancel: But wait, there’s more, such as a $5.2 million contract awarded by Defence, and $121,000 by the Department of Veterans’ Affairs, as reported by the Financial Review’s Edmund Tadros. The former was allowed because the procurement process happened before the ban while the latter was for services not covered by the ban.

Get outta my way: “Finance is not aware of any breaches of the agreement with KPMG that it will not bid for Australian Government contracts between 16 June and 31 October 2026,” a spokesperson told Playbook.

WHYALLA WIPEOUT — $5M CASH BACK: Sanjeev Gupta’s Whyalla steelworks is in the news again for cutting more than 500 jobs as administrator KordaMentha gets an extension to complete the sale of the distressed asset.

POLITICO can reveal OneSteel Manufacturing has also had to repay more than $5 million to the Australian government. Originally granted $63.2 million as part of the Powering the Regions Fund, only $10 million was paid out before the steelworks went into administration in February 2025. More than half that money has been returned, the rest was already spent in line with program rules.

AGENDA ASSISTANT

All the lovers: David Pocock is setting up shop at Gang Gang in Downer in the morning to talk to local constituents.

LOCAL CANBERRA

YOU BEE CRAZY!: Home to the Australian Government Solicitor, Productivity Commission and AUSTRAC, 4 National Circuit also counts several beehives as residents on the roof.

Where the wild roses grow: This week, it’s harvest time, and they’ve been selling the exclusive honey in the lobby (all proceeds donated to the Wheen Bee Foundation).

Today is your last chance to meet beekeeper Cormac Farrell. If you work in the building, his final tour starts at 12:15 p.m. from the lobby.

Cormac also looks after the bees at Parliament House. Its famous native stingless bees will be returning to Canberra next month after their annual winter holiday in Sydney.

LET’S GO FLY A KITE: Canberra’s annual kite festival kicks off — flies off? — at the National Arboretum at 10 a.m. Saturday. Bring the kids, or be one. You’ll get a second chance to see it from the same time Sunday.

After that: The Asia-Australia Youth Association’s Lake Moon Festival — complete with food, a water lantern show and cultural performances — begins at the Beijing Garden at 10 a.m.

Come into my world of medieval wonder: History buffs are grabbing their shields on Saturday and Sunday for Queanbeyan’s Medieval Festival, where you can witness a buhurt tournament. Google tells us that’s: “A full-contact combat sport where fighters wear steel armor and use blunt medieval weapons to knock opponents to the ground.” Gates open from 10 a.m.

YOUR DISCO NEEDS YOU: Get dancing — Substrate’s Inaugural Vinyl Night is happening Saturday.

POLITICO TRIVIA

What do I have to do, to get the message through — that you’ve been reading Kylie song titles throughout this Playbook, in honor the pop princess’ AFL Grand Final performance Saturday.

Can you identify how many references we snuck in?

Everyone with the correct answer gets a fancy POLITICO water bottle. Send the answer to [email protected].

COMMUNITY CORNER

MOVES AND APPOINTMENTS

Former Infrastructure Secretary Mike Mrdak has been reappointed as a non-executive director of the NBN Co. board for a three-year term starting Oct. 1.

BIRTHDAYS

TODAY: Disgraced former Victorian Labor MP Adem Somyurek, former South Australian Deputy Premier Kevin Foley, Western Australian Labor MP Lisa O’Malley, former U.S. Defense Secretary Robert Gates and Hollywood power couple Michael Douglas and Catherine Zeta-Jones both celebrate their birthdays today.

SATURDAY: Assistant Minister Andrew Charlton, New South Wales Labor MP Peter Primrose, Minister for the Pacific Pat Conroy’s adviser Claudine Lyons, former Ukrainian President Petro Poroshenko and tennis icon Serena Williams.

SUNDAY: Former W.A. Premier Geoff Gallop — who shares a birthday with his immediate predecessor Richard Court, shock jock Ray Hadley, N.S.W. Liberal MP Ray Williams and Oscar-winner Gwyneth Paltrow.

Home

OpenAI’s agents went rogue — its human response caused the real damage