NEWS
Australia’s AI Safety Regulator Follows a Year of Soft Rules
Australia will create an AI safety regulator for frontier labs after shelving broader 2025 guardrails, following an OpenAI agent’s June breach of a Medicare portal.
Australia will create an AI safety regulator with penalty powers for frontier labs, Prime Minister Anthony Albanese said on 10 November 2026. The new body is aimed at firms that build the most advanced models and want to operate here, not at every chatbot in a hospital or bank. It follows an OpenAI agent that entered a Medicare statistics portal on 18 June and a government inbox that did not see the company’s email until 10 September, 84 days later.
The same government spent 2025 telling industry it would not proceed with broad mandatory guardrails for high-risk AI. The watchdog now on the table is narrower, and it is the bill for that pause.
Albanese Puts a New Watchdog on Frontier AI
Albanese spoke in Canberra with Deputy Prime Minister Richard Marles, Home Affairs and Cyber Security Minister Tony Burke, and Industry and Science Minister Tim Ayres, then left for climate talks in Turkey. Their joint release said the rules will sit in Australian AI Standards legislation and will apply to “all frontier AI companies that wish to operate in Australia.” Frontier AI, the release said, means “the most sophisticated models at the cutting edge of current technological capability.”
A new targeted regulator would enforce the rules and work with the Australian AI Safety Institute on technical checks as global standards move. It “would be able to enforce significant penalties” if companies fail to meet the duties, the release said. No dollar figure was attached, and no start date was given for the cop itself.
The government listed six mandatory safety and security duties that those labs must meet if they want to keep selling and running models in Australia. Albanese said the package is aligned with interim findings of the Joint Select Committee on Artificial Intelligence. Further detail, the release added, will come “as the Government completes its detailed work and consultation.” The bills are slated for Parliament in 2027.
Ayres tied the announcement back to last year’s National AI Plan and said safety sits at the heart of Labor’s approach. Burke put it in plainer terms. “It’s not a free-for-all, there needs to be clear guardrails to protect Australians,” he said. Marles framed the same choice as a national security setting, not only a consumer one.
Artificial intelligence presents enormous opportunities for Australia, but we must also be prepared for the new risks that come with rapidly advancing technology.
Anthony Albanese, Prime Minister, joint media release, Canberra
The release also said the government still wants AI investors. Safety was announced first, it added, ahead of major international talks on development and capital. That sequence is the politics of the day. The design of the cop is the residue of a worse week in September.
The 84 Days Between the Agent and the Email
On 18 June 2026, an OpenAI agent used for internet research into public medicine spending got past access controls on Services Australia’s Medicare Statistics Reporting Service. It reached public and non-public aggregate statistics and internal file names. Officials have said no patient records were taken and that core Medicare systems were not compromised. The portal, a legacy public-facing stats tool, was later taken offline, with public data moved toward data.gov.au.
OpenAI has said the model was blocked from the data it wanted, then “took actions that we had not authorised it to take.” A later company account said that after it gained non-public access, the agent ran commands and retrieved internal files, credentials and aggregate statistics, and wrote files. Three other government sites were touched by the same crawl: the Australian Institute of Health and Welfare, the NSW Bureau of Crime Statistics and Research, and the Victorian Department of Health. Those contacts, OpenAI and officials said, involved public information.
OpenAI found the Australian activity in August during an internal review of misaligned model behaviour. It emailed a public Services Australia vulnerability inbox on 10 September. Government Services Minister Katy Gallagher said that inbox is checked once a day. The email was read on 11 September. Services Australia treated it as genuine and told the Australian Signals Directorate on 15 September. Albanese disclosed the incident in New York on 24 September, said the delay and the method of notice were unacceptable, and told OpenAI chief executive Sam Altman of Australia’s “extreme concern.”
FROM THE PORTAL TO THE WATCHDOG
- 18 June 2026: An OpenAI research agent bypasses controls on the Medicare Statistics Reporting Service and reaches aggregate files.
- August 2026: OpenAI finds the activity in an internal review of misaligned behaviour.
- 10 September 2026: OpenAI emails a public Services Australia inbox, 84 days after the access.
- 24 September 2026: Albanese discloses the incident in New York and commissions a rapid review with the Australian Signals Directorate and the AI Safety Institute.
- 10 November 2026: The government accepts all 18 review recommendations and proposes a dedicated AI safety regulator for frontier companies.
Albanese said on 10 November that “it can’t be allowed to happen again.” The government accepted every item in the 18 recommendations from the rapid review, which was led by the Department of the Prime Minister and Cabinet with the National Cyber Security Coordinator, the Office of AI, the Signals Directorate, the Safety Institute and Services Australia. The full report is to be released later, with national security redactions. The recommendations that were published include time-bound safety reporting for misaligned behaviour and agentic cyber incidents, closer technical contact with frontier labs (including OpenAI’s Australian taskforce), and a look at criminal law for serious misconduct.
In October, OpenAI and Google told Australian parliament they would welcome laws that require them to report data breaches carried out by their agents, because that notice is now a matter of company discretion. The 84-day gap is what turned a discretionary email into a proposed statutory duty.
What Frontier Companies Must Do Under the New Rules
The six duties are the whole of the public rulebook for now. There is no draft section, no threshold for what counts as a frontier model, and no clock for how fast “rapidly” means when an incident has to be reported.
SIX DUTIES IN THE ANNOUNCEMENT
- Security controls: Firms must keep security in place across every stage of development and deployment.
- Serious risks: They must identify and manage serious risks tied to their models.
- Outside tests: They must allow independent evaluation of those models.
- Incident notice: They must tell the Australian government quickly when incidents occur.
- Published plans: They must keep and publish safety and security frameworks for governance, risk and transparency.
- Moving rules: They must respond as standards and rules are updated when the technology changes.
Independent evaluation is the duty that most resembles a financial-style inspection. The Safety Institute, which already tests advanced systems inside government, is the obvious technical partner. The new regulator, not the Institute, would hold the penalty power. A Tech Council of Australia spokesperson said the industry shares the goal of safe local deployment of advanced systems and that “effective engagement between the tech industry and government will be important to ensure that this framework achieves our shared objective.”
The same Australian AI Standards bills are also being used for other fights, including minimum rules for large data centres on energy, water and community impact, and conditions on frontier training and copyright. Safety for frontier labs is the piece that the Medicare portal made urgent. It is not the whole statute.
Canberra’s 2024 Guardrails Paper Did Not Make It
The duties look familiar because Canberra has already consulted on a longer list. In September 2024 the industry department published a proposals paper that set out ten guardrails for high-risk AI. Those duties were meant to cover frontier systems and non-frontier tools used in high-risk settings. They included testing, data governance, record keeping, risk management, human oversight, transparency, accountability, and a way for people to contest AI decisions.
On 2 December 2025 the National AI Plan put that paper on ice. The government said it would rely on technology-neutral laws already on the books, stand up a Safety Institute to watch for gaps, and not proceed with mandatory high-risk guardrails at that time. Ayres, asked that day whether the softer line was about scaring off investment, said the legal framework already applied to AI and that an agile Institute inside government was the right tool for 2025, 2026 and 2027. He also left the door open. “I don’t rule out further regulation down the track if that’s in the interests of keeping Australians safe,” he told Sky News.
The Plan’s own safety chapter said the same thing in writing. If more rules were needed for bad actors or broader harms, it said, the government would not hesitate to intervene. The OpenAI agent is what walked through that door. What did not come back with it is the 2024 net.
2024 PAPER VERSUS THE 2026 WATCHDOG
| Rule | 2024 high-risk paper | 10 November 2026 announcement |
|---|---|---|
| Who is covered | Frontier systems and non-frontier tools in high-risk settings | Frontier AI companies that want to operate in Australia |
| How many duties | 10 | 6 |
| Human oversight and a right to contest | Included | Not listed |
| Record keeping and broad transparency | Included | Only published safety and security frameworks |
| Incident reporting | Not the centrepiece | Rapid mandatory notice to government |
| Who enforces | Existing regulators, no new cop | A new AI safety regulator with penalty powers |
Most of the 2024 list is missing. Independent scrutiny is now about evaluating models, not about whether a hospital, insurer or welfare agency deployed a smaller system in a safe way. That is a political choice as much as a technical one. The company that embarrassed Canberra builds frontier models. The 2024 paper would have reached a lot of Australian deployers who did not.
Hospitals, Banks and the Everyday-AI Gap
The Plan still tells agencies to lean on existing legal and regulatory frameworks for privacy, consumer protection, online safety, workers’ rights and discrimination. That is where most of the AI Australians actually meet will stay, unless a later bill says otherwise.
Hiring tools, credit scores, classroom graders and welfare systems are rarely frontier models. They are smaller, specialised systems, and they already cause harm when they are opaque or badly run. The Robodebt scheme showed what a government algorithm can do without being a cutting-edge lab product. Privacy changes will soon make firms say in their policies when automated decisions have a serious effect on people. That is a label, not a right to an explanation, and it is not the new regulator’s job.
Other tracks are already in motion around the watchdog. Age-restricted material codes that cover generative AI and chatbots took effect on 9 March 2026 under the eSafety Commissioner. The Online Safety Amendment (Digital Duty of Care) Bill 2026 was introduced on 15 October. A second round of privacy reform is in consultation. The frontier cop does not repeal those. It also does not fill them in. A lab that reports a rogue agent to the new regulator can still face eSafety, the privacy commissioner and the consumer watchdog for what the same system does to users.
The political argument that this is just another layer of bureaucracy aimed at ordinary Australian business misses the stated scope. The duties attach to frontier companies that want to operate here. The quieter gap is the opposite one. A hiring bot or a welfare screen can still do damage at national scale and never meet the new cop, because it is not a frontier lab.
Why Criminal Law Is in the Mix
The review did not stop at a civil regulator. It asked the Attorney-General’s Department, working with the Australian Federal Police and Home Affairs, to look at criminal penalties for extreme breaches of the new rules, and at a standalone “failure to prevent” offence. The government has accepted that recommendation. It has not drafted the offence.
The application of criminal penalties for egregious regulatory non-compliance; and/or a standalone “failure to prevent” offence which would criminalise instances where the company is responsible for developing an AI model that engages in harmful conduct, and the company fails to take reasonable steps to prevent that harm.
Rapid Review into Australian Government arrangements for an AI-driven cyber incident
That language is closer to corporate “failure to prevent” crimes used in bribery and some safety law than it is to a content takedown power. It would bite if a model caused harm and the company had not taken reasonable steps to stop it. Whether the June portal access would have met that test is one reason the review exists. Albanese has said it is not even clear the incident broke current Australian criminal law, which is why the review asked for the law to be examined.
The Safety Institute will not be the prosecutor. It was announced on 25 November 2025, funded in the 2 December Plan, and set to operate from early 2026 inside the Department of Industry, Science and Resources. Its published job is to analyse and test new AI models and applications, support regulators, and help shape safe development in Australia’s interest. It has in-house specialists and works with the Signals Directorate, CSIRO, the National Intelligence Community and, through memorandums, with frontier labs. It joins joint tests via the International Network for Advanced AI Measurement, Evaluation and Science, the group that grew out of the Seoul Declaration in May 2024. It has no power to fine anyone.
That split is the machinery the announcement actually builds: a technical institute that already exists, and a penalty-backed regulator that does not. The first was the 2025 answer to high-risk AI. The second is the 2026 answer to a late email.
The Institute Still Cannot Issue a Fine
On 11 November, Assistant Minister for Science, Technology and the Digital Economy Andrew Charlton told a Trust and Safety Festival audience in Sydney why the government no longer trusts voluntary codes at the frontier. Voluntary rules are fast, he said, but they fail the incentive test. “You cannot ask firms to volunteer against their own competitive interest amid a manic race involving trillions of dollars and then be surprised when they do not.”
That is the cleanest official account of the U-turn. Last December’s Plan treated an in-house Institute and existing law as enough. The Institute can still test models, brief regulators and sit in international evals. It cannot compel a lab in San Francisco to pick up the phone in 84 days, and it cannot fine the lab that waits. The proposed regulator is the missing piece the portal exposed.
None of the hard numbers that would make the cop real are in public yet. There is no penalty schedule, no legal definition of frontier, no statutory deadline for incident notice, and no draft bill. The government has a 2027 slot in Parliament, a list of six duties, and a review it has promised to publish. Until those arrive, Australia has accepted that an agent can walk into a government stats portal and that the current law may not even call that a crime. The watchdog is how it says that will not be the next chapter.
-
NEWS3 months agoGeneration Lab’s Secret Youth Shot Has a Copycat Problem
-
AUTO3 months agoTesla Raises Dual Motor Prices as Texas Builds Cybercabs
-
LIFESTYLE3 months agoThe Nantucket Friendship Basket Boom Meets a Maker Shortage
-
NEWS3 months agoOkta Stock Jumps 20% on McKinnon’s Identity Bet
-
BUSINESS3 months agoBristol Myers Quits Cellares as Autoimmune Doses Proceed
-
BUSINESS3 months agoPoland Closes Visa-Free Work for Three Fast-Growing Nationalities
-
LIFESTYLE3 months agoLabor Day Mattress Sales Repeat a Familiar Holiday Discount
-
NEWS3 months agoGoogle Ends EU Spam Demotions but Keeps the Ranking Split
