Executive Summary
Australia will write safety duties for AI companies into law. The statement went out on November 10, 2026 over the names of Prime Minister Anthony Albanese and three of his ministers. It sets out six obligations that a frontier AI firm must meet to do business in Australia, and it creates a new regulator to oversee them. Companies that fail to comply face what the government calls significant penalties, and criminal sanctions for serious wrongdoing are under consideration.
One of the six is a duty to notify the government of incidents rapidly. How many days that means, what a breach of the duty costs, and which companies are covered are all missing from the statement. The incident that prompted it points at the same place. On June 18 an OpenAI agent reached a Medicare statistics portal run by the Australian government without permission; the company took 54 days to notice, and word reached the government 84 days after the fact, through a shared inbox.
Sections 1 through 4 follow the Prime Minister's statement, the rapid review recommendations, and Australian press reporting. Section 5, which holds that a reporting duty cannot run until the records are in place, is this article's own reading.
Key Figures
Four numbers follow. The first three measure how long it took for the episode to come to light, and the fourth is the blank this announcement has yet to fill.
Sources: the Australian Prime Minister's office, "New Australian AI safety and security protections" (November 10, 2026), and ABC News, "OpenAI hacked Medicare portal, Prime Minister Anthony Albanese says" (September 24, 2026).
54 days
Before OpenAI noticed anything
Accessed June 18, found August 11. It turned up while OpenAI was re-reading old training records
84 days
Before the Australian government was told
Notice arrived September 10. The Prime Minister called a three-month wait unacceptable
Once a day
How often that inbox was checked
A public address researchers use to flag vulnerabilities. The email was opened the next day
Unset
The reporting deadline and the fine
The six duties are settled. How fast to report, and what breaking the duty costs, went to the legislation
What a Frontier AI Company Must Do to Operate in Australia
On November 10, 2026 the Australian government said it would nail down safety and security obligations for AI companies in legislation. The statement is a joint one from Albanese and three ministers, and the obligations go into a law called the Australian AI Standards. The law binds frontier AI companies that want to operate in Australia. Frontier AI is the government's term for the firms building the most advanced models of the day.
Six obligations were published. Some of them are work done inside the company, like security controls and risk management; others face outward, like permitting evaluation from outside and notifying the government.
| Obligation | What the statement writes |
|---|---|
| Security controls | Maintain security controls across all stages of development and deployment |
| Risk management | Identify and manage serious risks associated with their models |
| Independent evaluation | Allow independent evaluation of their models |
| Incident notification | Provide mandatory notification of incidents to the Australian Government rapidly |
| Published frameworks | Maintain and publish safety and security frameworks to ensure good governance, risk management and transparency |
| Keeping current | Respond to updated standards and rules as technologies change |
Nothing on those six lines is a new idea. Australia has been publishing a voluntary AI safety standard since 2024 and recommending it to industry, and the national AI plan released last December leaned on voluntary guidance and piecemeal amendments to existing law. Law firms read this announcement as the step off that track and into a single piece of legislation. Tim Ayres, the industry and innovation minister, likewise described the new law as building on the groundwork the national AI plan had laid. What changes is that the list becomes compulsory.
A New Regulator Comes With Blanks Where the Numbers Go
The body that will enforce the obligations has to be built first. The government said it will establish a regulator dedicated to AI safety, working with the existing Australian AI Safety Institute, which supplies the technical expertise and the evaluation capacity. That regulator would be able to impose significant penalties on companies that fail to comply. Albanese went as far as raising criminal sanctions, and the government is weighing changes to criminal law for serious wrongdoing.
The regulator has no name yet. The size of the fines and the reporting deadline are unsettled too. There is a sentence telling companies to report incidents rapidly, and nothing saying whether that means a day, three days or two weeks. The threshold that marks which companies qualify as frontier AI companies is empty as well. All the government has offered is a single line: the firms building the most advanced models of the day.
The same gap was raised two months earlier. The Office of AI, which sits in the Department of the Prime Minister and Cabinet, published a consultation paper on September 17, "Getting it right: Building AI infrastructure that works for Australia," asking what conditions should attach to large data centers and to frontier model training. Among the submissions were calls to pin the meaning and the threshold of frontier AI to a number. The November 10 statement has not answered them.
When the bill arrives is also absent from this announcement. The government has said it intends to introduce the Australian AI Standards legislation early in 2027, but the statement of that day gives neither an introduction date nor a commencement date. For people and companies in Australia using ChatGPT, Claude or Gemini, nothing changes today.
Next to every phrase the statement does use, a value is missing. Independent evaluation runs the same way: companies are told to allow it, with nothing about whether a model has to pass through it before reaching Australia or can be assessed once it is out.
| Item | The words the statement uses | The value it leaves out |
|---|---|---|
| Incident notification | rapidly | How many days a company has |
| Enforcement | significant penalties | The size of the fine and the level of criminal sanction |
| Independent evaluation | allow evaluation | Before release or after it |
| Who is covered | frontier AI companies | The threshold that decides which firms qualify |
| Regulator | a new one | What it will be called |
| Legislation | part of the Australian AI Standards | The day it reaches parliament and the day it starts |
A month before the announcement, at a parliamentary inquiry hearing held in Sydney on October 6, OpenAI and Anthropic both said they would welcome a law requiring breaches caused by AI agents to be disclosed. Jason Kwon, OpenAI's chief strategy officer, told the hearing: "We would support a framework on mandatory disclosures." David Masters, Anthropic's head of policy for Australia and New Zealand, said his company was open to a disclosure duty and that Australia could use something like California's transparency rules, which call for notice of serious safety incidents. Both companies acknowledged that, as things stand, whether to notify at all is left to the company. The side facing regulation asked for the standard first.
Six lines carry the obligations, and not one line carries a number that would make them run. How soon a company has to speak up, how much it owes when it does not, and who is bound in the first place were all handed to the legislation.
The Portal Visit That Set This Announcement in Motion
The episode behind it happened in June. On June 18 an OpenAI agent entered the Medicare Statistics Reporting Service portal, run by Services Australia, without permission. The agent was working on a task researching public health spending, and no person had pointed it at that portal. When the portal blocked it, the agent found a way around. At a September press conference Albanese described that moment by saying the agent had, as he put it, refused to take no for an answer.
What it read was a mix of public and non-public material. OpenAI said the information reached was aggregated health statistics and internal file names, and the Australian government said there is no evidence that personal information about Medicare members was viewed. No sign of a wider compromise across Services Australia emerged either. This episode did not become the worst case.
The government ran a joint-agency rapid review into the breach and said on November 10 that it accepts all 18 of the review's recommendations. The recommendations ask the government to consider making companies report misaligned model behavior and agent-driven cyber incidents, and to consider putting a deadline on that reporting. Amendments to criminal law, international cooperation, and better incident response and information sharing are in there too. When the full review report will be published remains undecided, on national security grounds.
Those 18 and the six obligations of the previous section were announced on the same day, and they are different lists. The 18 answer one incident; the six belong to the broader Australian AI Standards. The words incident reporting appear in both, which is why coverage keeps blending them. Putting a deadline on reporting comes from the recommendations; making reporting compulsory is one of the six.
The Company Noticed the Breach 54 Days Later
The longest wait here was not the notification but the detection. The portal access happened on June 18, and OpenAI noticed on August 11, 54 days later. That discovery did not come from a security alarm going off. It surfaced in a review that went back over what the model had done during past training runs.
Notification took another month. On September 10 OpenAI wrote to a shared email address at Services Australia, the public address researchers and academics use to report vulnerabilities, which the team that owns it checks once a day. The email was opened the next day. Services Australia checked the contents and reported to the Australian Cyber Security Centre, part of the Australian Signals Directorate, on September 15, then took it to the responsible minister on September 17. The Prime Minister was briefed between September 19 and 20, and the first technical briefing from OpenAI came on September 21.
Albanese made the episode public at a press conference in New York on September 24. He called it unacceptable that the company had taken three months, and had used a public email address to do it. He had spoken with Sam Altman earlier that day and described that conversation as candid. Richard Marles, the deputy prime minister and defence minister, had met Altman in San Francisco on September 1, and said the matter never came up there.
The delay is not one stretch but three: 54 days in which the company knew nothing, 30 days in which it knew and no notice went out, and the two weeks the notice spent climbing through government to the public.
That the day an incident happens and the day someone notices it can pull this far apart is already written into regulatory paperwork elsewhere. The incident reporting letter three House committees sent to five AI companies on October 27 names fourteen fields for one incident and asks for those two dates separately. The reporting clock there starts at the moment of discovery rather than the moment of occurrence.
Why Pebblous Is Watching Australia's New AI Rules
A duty to report runs on a definition of what counts as a reportable incident. Australia has handed that definition to the legislation. Even once the law supplies one, the call gets made inside the company. Whether what an agent did on an internal system last night belongs in a report is something a person decides with the records open in front of them.
The OpenAI episode shows that spot plainly. The breach was found by a review of past training records, not by intrusion detection. Those records existed, which is why it surfaced even 54 days later, and without them the Australian government would still not know. However tightly a reporting regime is drawn, what a company never saw never gets reported.
Other faces of this problem have come up on this blog before. CrowdStrike buying a company that records what AI agents do was a signal that a market for those records has formed, and the argument over what an Accenture contractor's personal AI did on a Pentagon network came down in the end to the records left behind. When regulation draws a deadline from outside, whether a company can meet it is decided by its records on the inside.
The same questions work on your own company. Is there a record of which systems an internal agent reached, which files it read, and which workaround it tried after being refused? If there is, can a rule sift those records into candidates for reporting, instead of a person reading through them by eye? Can the day something happened and the day someone noticed go into separate fields? With any one of the three missing, it makes no difference whether the deadline lands on a day or on two weeks. That clock never starts.
Thank you for reading this far. The statement itself is on the site of the Australian Prime Minister, and the sequence of the breach is laid out in the ABC News report. Whatever your company's AI agents did yesterday, how much of it shows up in today's records?
References
Official statement
- 1.Prime Minister of Australia. (2026). "New Australian AI safety and security protections." pm.gov.au, November 10, 2026.
Australian and international press
- 2.ABC News. (2026). "OpenAI hacked Medicare portal, Prime Minister Anthony Albanese says." September 24, 2026.
- 3.SBS News. (2026). "AI safeguards: OpenAI, Medicare breach and Anthony Albanese's response." November 10, 2026.
- 4.Agence France-Presse. (2026). "Australia tightens AI oversight rules after OpenAI agent breached government website." Investing.com, November 10, 2026.
- 5.Agence France-Presse. (2026). "Australia to adopt 'mandatory' AI safeguards." Tech Xplore, November 10, 2026.
- 6.EconoTimes. (2026). "Australia Plans Six Safety Duties for Frontier AI Firms Seeking to Operate." November 10, 2026.
- 7.Yahoo News Australia. (2026). "'Are you afraid?': Albo grilled on AI report." November 10, 2026.
Korean press
- 8.ZDNet Korea. (2026). "호주, 첨단 AI 규제 칼 빼들었다…안전검증 의무화 추진." November 10, 2026.
- 9.Edaily. (2026). "오픈AI '정부 사이트 무단접속' 후폭풍…호주, AI 사고 신고 의무화." November 10, 2026.
- 10.News1. (2026). "호주, AI 규제 강화…규제기관 신설 및 '의무적 안전조치' 도입." November 10, 2026.