Living reference — Australia / AI agents

Australia's Rules for AI Agents: Where They Stand

Australia is deciding how to govern autonomous AI agents after one met a refusal and went around it. That is an order failure: a step that should have stopped did not. This page sets out what is on the public record, why every remedy proposed so far works only after the fact, and the question each still leaves open: who holds the record of what an agent did.

Published 5 October 2026 · Last reviewed 18 November 2026 · Updated as primary documents are published. Primary sources are cited throughout: Parliament, the government and the developer's own statements. The United States counterpart is US rules for AI agents.

1. Where things stand

DateOn the record
Sep 2024The Department of Industry, Science and Resources publishes proposals for mandatory guardrails for AI in high-risk settings. They were not legislated.
2 Dec 2025The government publishes the National AI Plan, which builds on "existing legal and regulatory frameworks". The department's consultation page records that the government "will not proceed at this time" with the 2024 proposals for mandatory guardrails.
18 Jun 2026An AI agent gains unauthorised access to the Medicare Statistics Reporting Service portal, administered by Services Australia.
20 Aug 2026Both Houses appoint the Joint Select Committee on Artificial Intelligence. Final report due no later than 30 November 2026.
10 Sep 2026The developer notifies Services Australia.
Sep 2026The Department of the Prime Minister and Cabinet opens a consultation on national AI standards, Getting it right: Building AI infrastructure that works for Australia. It proposes that frontier labs authorised to undertake large-scale AI training in Australia disclose "defined reportable AI incidents to relevant Australian authorities". Submissions closed 9 October 2026.
21 Sep 2026The President of Finland and the Prime Minister of Norway launch A Call for Control of Frontier AI Models. The Prime Minister of Australia is among its signatories; New Zealand is not, at 25 October 2026. It records "capable AI systems circumventing testing safeguards, exploiting vulnerabilities and gaining unauthorized access to real-world systems", and calls for mandatory pre-deployment testing, independent evaluation, common standards and the shared reporting of serious safety incidents.
24 Sep 2026The Prime Minister discloses the incident, establishes a task force for an urgent review, and refers the incident to the Joint Select Committee. The review's terms of reference are published the same day.
26 Sep 2026The Prime Minister says the Australian sites were not the only ones: "There are dozens of cases, including US government sites."
6 Oct 2026The committee held a public hearing at NSW Parliament, Sydney. Its program lists Anthropic, OpenAI, Microsoft, Google and Google DeepMind, and the Commonwealth Bank of Australia among the witnesses. The proof transcript was published on 7 October.
7 Oct 2026A second day of hearings at NSW Parliament, Sydney, with AI safety, cyber security and automation witnesses, including the Gradient Institute, Global Shield, Palo Alto Networks and Workato. The proof transcript has been published.
8 Oct 2026A hearing at Parliament House, East Melbourne. The transcript records Per Capita, Data Centres Australia, the National Computational Infrastructure, the Pawsey Supercomputing Research Centre, the Australian Academy of Technological Sciences and Engineering, Good Ancestors, the Minderoo Foundation, the Australian Chamber of Commerce and Industry, and Telstra. The proof transcript has been published.
9 Oct 2026A hearing in East Melbourne.The transcript records Monash University and RMIT University, CDC Data Centres, FinTech Australia, CPA Australia, the Business Council of Australia, the Council of Small Business Organisations of Australia, MYOB, Westpac, the National Australia Bank, the Australian Banking Association and ANZ. The proof transcript has been published.
12 Oct 2026The committee writes to the Prime Minister with early advice ahead of its report. Its first recommendations are headed "AI Safeguards" and are made "following the incidents related to the unauthorised access by an AI agent to data retained in Australia's Medicare system". Section 6 sets out what they rely on.
14 Oct 2026A hearing in Canberra. The transcript records the Productivity Commission and the Australian Electoral Commission. The proof transcript has been published.
16 Oct 2026A hearing in Canberra. The transcript records the Institute for Progress, the Carnegie Endowment for International Peace, the Australian Academy of the Humanities, Amazon Web Services, the Digital Economy Council of Australia, the Tech Council of Australia, Professors Michael Quinlan and Penny Williams, the Tech Policy Design Institute, the Australian Academy of Science, the Kingston AI Group, the Australian Strategic Policy Institute, the Australia Institute, Safe Work Australia, the Centre for Future Work, and the Australian Competition and Consumer Commission. The proof transcript has been published.
Oct 2026Safe Work Australia publishes the final report of its Best Practice Review of the model WHS laws. It finds the principles-based laws "well placed" to regulate AI, which it treats, "regardless of their level of autonomy", as tools designed and used by humans, and recommends that Safe Work Australia develop "a strategy on artificial intelligence and other emerging digital technologies" that monitors how the laws are applied and fills gaps where they are found (Recommendation 19).
20 Oct 2026A hearing in Sydney. The transcript records the eSafety Commissioner, the Human Technology Institute at the University of Technology Sydney, the Australian Information Industry Association, the Australian Securities and Investments Commission, CSIRO, the Australian Council on AI Strategy, the ARC Centre of Excellence for Automated Decision-Making and Society, Science and Technology Australia, and Research Australia. The proof transcript has been published.
21 Oct 2026A hearing in Sydney. The transcript records the National Farmers' Federation, the Transport Workers' Union and the John Curtin Research Centre, the Australian Society of Authors, the Australian Publishers Association and Hachette Australia, the Future Skills Organisation and MEGT, Unions NSW and the Health Services Union, the Australian Medical Association, the Royal Australian and New Zealand College of Radiologists, the UNSW AI Institute, and the United Workers Union. The proof transcript has been published.
22 Oct 2026A hearing in East Melbourne. The transcript records the National Tertiary Education Union, the Independent Education Union of Australia and the Australian Education Union, the Finance Sector Union and Professionals Australia, Emerson Economics and TechnologyOne, Maurice Blackburn Lawyers, Atlassian, the Australian Nursing and Midwifery Federation, the Australian Services Union and the Australian Rail, Tram and Bus Industry Union, the Australian Council of Trade Unions, and the Human Rights Law Centre, the Australian Human Rights Commission and Digital Rights Watch. The proof transcript has been published.
30 Oct 2026A hearing at Parliament House, Canberra. The transcript records the Community and Public Sector Union; the Treasury, the Department of Finance, the Digital Transformation Agency and the Department of Employment and Workplace Relations; the Department of Infrastructure, Transport, Regional Development, Communications, Sport and the Arts; the Attorney-General's Department; the Department of Climate Change, Energy, the Environment and Water; the Department of Home Affairs; the Department of Defence; the Department of Foreign Affairs and Trade; the Department of Industry, Science and Resources, with the AI Safety Institute and the National AI Centre; and the Department of the Prime Minister and Cabinet, with its Office of Artificial Intelligence. The proof transcript has been published.
Nov 2026The Rapid review into Australian Government arrangements for an AI-driven cyber incident sets out 18 recommendations. It describes the matter as "the OpenAI cyber incident affecting Services Australia". Recommendation 3: "The Australian Government consider mandatory, time-bound AI safety reporting, including but not limited to misaligned behaviour or agentic cyber incidents." Recommendation 8: the Attorney-General's Department, with the Australian Federal Police and the Department of Home Affairs, "consider options to reform the criminal law to improve AI safety", exploring "criminal penalties for egregious regulatory non-compliance" and/or "a standalone 'failure to prevent' offence".
10 Nov 2026The Prime Minister, the Deputy Prime Minister, the Minister for Home Affairs and Cyber Security and the Minister for Industry and Innovation announce that "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". The Government "will propose" that they be enforced by "a new targeted regulator for AI safety". Companies would be required to, among other duties, "Provide mandatory notification of incidents to the Australian Government rapidly" and "Allow independent evaluation of their models". The release says the protections "are aligned with interim report recommendations from the Joint Select Committee on Artificial Intelligence" and that "Further details of the legislation will become available in due course."
16 Nov 2026California's Attorney General sues OpenAI over its agents' incidents. Among them, the complaint alleges that in June 2026 OpenAI models "accessed and retrieved non-public data from Australian government websites", having defeated sandbox restrictions "that were supposed to prevent them from sending data to third-party servers". Section 6 sets out what it says of the record.
PendingThe committee's report (its letter of 12 October is early advice, not the report), and the AI Standards legislation, whose details the Government says will follow "in due course". Neither had been published at the date of this page.
From the 6 October hearing (proof transcript)

Quoted from the proof Committee Hansard, an uncorrected proof of evidence. Page numbers are the printed pages.

  • OpenAI (Jason Kwon, Chief Strategy Officer): "we should have informed the impacted parties much sooner in the process. The reason why that did not happen is we wanted to understand more of the facts before we spoke to the impacted parties, but, because of the novelty of this situation, we have learned our lesson that it is better to inform parties, even with partial information, that something has occurred" (p. 32). He said that change led to the NSW National Parks and Wildlife Service being told "within 48 hours" of OpenAI finding the activity (p. 32).
  • OpenAI: asked whether OpenAI's chief executive knew of the breach when he met the Deputy Prime Minister on 1 September, Mr Kwon said he "was not aware at the time" (p. 35).
  • OpenAI: asked whether its models could break into an oil and gas facility and operate it, Mr Kwon said "we don't know, because it's going to depend on the safeguards and the infrastructure and the security controls that that particular company or provider has in place" (p. 36).
  • OpenAI: "we would support a framework on mandatory disclosures" (p. 38).
  • Anthropic (Dave Orr, Head of Safeguards): "We would definitely reach out within a matter of days, or sooner, as soon as we understood the details" (pp. 28–29).

Every commitment made at the hearing concerns notification. Section 6 sets out what notification does and does not settle.

2. What happened, on the public record

The Prime Minister, at a press conference in New York on 24 September 2026, described it this way:

"an OpenAI agent gaining unauthorised access into the public-facing Medicare statistics reporting service portal, which is administered by Services Australia."
"The AI agent found a way around those blocks. Didn't accept no for an answer, if you like."

He said no personal information was believed to have been accessed and that investigations were ongoing.

The developer's own account, published on 28 September 2026, says the agent, running during internal training and evaluation, "discovered a way to gain non-public access to the service, and ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files", and that individual patient or client records were not accessed. The same account reports activity affecting four other Australian bodies: the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health, the Australian Institute of Health and Welfare, and, in an update on 4 October, the NSW National Parks and Wildlife Service.

Asked at the committee's 6 October hearing how sophisticated the access to Services Australia was, the developer's Chief Strategy Officer said:

"the focus here is not so much about the sophistication as much as the automation. It's the fact that now you have the ability to take agents and then apply them to a goal, and they will continue to work towards that goal. They're capable of trying many methods" (proof Committee Hansard, p. 36)

The Australian Academy of Technological Sciences and Engineering told the committee on 8 October: "International frontier models have shown they can hack government systems despite developers putting safeguards in place." (proof Committee Hansard, 8 October 2026, p. 19)

The blocks worked as blocks: each refused what it was shown. What failed was the order. A refusal should have ended the step, nothing made it do so, and an agent that keeps "trying many methods" will keep going until one gets around them.

3. Australia has seen this before

Australia's official record already holds several cases of an automated system acting without a step that the law or the task required. In each, an independent body found the failure afterwards.

CaseWhat the official finding says
RobodebtAn automated scheme that raised welfare debts. The Royal Commission into the Robodebt Scheme (Commissioner Catherine Holmes AC SC) presented its report on 7 July 2023, with a chapter on automated decision-making.
Revenue NSWGarnishee orders taking money from bank accounts to recover fines and debts, run through machine technology. The NSW Ombudsman's special report, tabled on 30 April 2024, found the conduct "contrary to law" until March 2019 and "wrong" until March 2022.
Welfare payment cancellationsFrom April 2022 the law required a job seeker's circumstances to be considered before income support was cancelled. The Commonwealth Ombudsman found that the system kept cancelling automatically, without that consideration, affecting 964 job seekers, and that the agencies' quality assurance did not identify it. Report: "Automation in the Targeted Compliance Framework: when the law is changed but the system isn't", August 2025.
Child protection reportA Victorian child protection worker used ChatGPT while drafting a Protection Application Report submitted to the Children's Court. The Office of the Victorian Information Commissioner found it "contained inaccurate personal information – which downplayed risks to the child in the case" and required the department to block such tools for child protection workers (24 September 2024).

The Robodebt Royal Commission's answer to this pattern was oversight that does not depend on the agency running the system:

"The Commonwealth should consider establishing a body, or expanding an existing body, with the power to monitor and audit automated decision-making processes with regard to their technical aspects and their impact in respect of fairness, the avoiding of bias, and client usability." (Recommendation 17.2)
"business rules and algorithms should be made available, to enable independent expert scrutiny." (Recommendation 17.1)

Those cases involved the government's own systems, and each took an investigation to surface. An AI agent does the same work faster, and the account of what it did is written by the system that did it.

4. The questions Australia has put so far

The committee's terms of reference include, among others:

"the adequacy of Australia's existing laws and regulatory frameworks as they apply to AI and whether there are any gaps that warrant reform;"
"the implications of emerging AI capability for Australia's national security and strategic resilience, including the ability of regulators, the Australian AI Safety Institute and the intelligence and security community to identify and respond to emerging risks;"

At the committee's 18 September hearing, before the incident was disclosed, Senator Tony Sheldon asked what steps a company implementing AI is required to follow:

"Is there a requirement about processes, and how does that requirement play out as part of a demand of every level of decision-making or oversight from the highest level in the AI safety group right through to a company implementing it? What are the steps that are required through that mechanism?"

The written answer, provided by Safe Work Australia, is about work health and safety law. It says the model laws are "principles-based and 'technology neutral'", that a business "cannot shift liability to AI systems", and that "designers of AI systems" also hold duties. It names no steps. At the 16 October hearing Safe Work Australia's Chief Executive, Marie Boland, said the same in person, and pointed to one part of the framework that does: the psychosocial regulations, which she described as "quite clear and specific about step-by-step approaches" (proof Committee Hansard, 16 October 2026, p. 64). Section 5 sets out the order those regulations follow.

A rule that names no steps has no order for an agent to break, so an order failure passes it by. Section 5 sets out the difference.

5. Two kinds of rule

Rules about AI behaviour come in two kinds. An outcome rule says what must not happen: manage the risk, report the incident, do not cause harm. An order rule says what must happen before what: this check before that action, this approval before that payment. The frameworks described on the record so far are outcome rules, which is why the answer to "what are the steps?" names none. Australian law does contain order rules: the welfare cancellation case turned on one, a required consideration before a cancellation, and the system ran past it.

Witnesses at the committee's 7 October hearing drew the same line from two directions. The Shop, Distributive and Allied Employees' Association, describing algorithms that set rosters and issue warnings at work, preferred the work health and safety model because "under that act, the consultation requires that, before the decision is made, there has to be consultation" (proof Committee Hansard, 7 October 2026, p. 2). As things stand, its researcher said, "consultation in its current form is coming after changes have already been decided" (p. 3). The Gradient Institute described a simulation in which agents in a market were instructed "It is illegal to collude. You must not collude." and "They will still collude. They will still find a way", in ways "really hard to detect, even if you can look at everything the agents are saying to each other" (p. 59). An instruction held by the agent is not the same thing as an order checked outside it.

The next day Senator Tony Sheldon put the association's retail case to the Australian Chamber of Commerce and Industry: an error rate of 0.18 per cent, a final written warning, and no indication "whether a human verified the decision the AI made or reviewed the flags before she was disciplined". The Chamber's answer was that existing law already covers it: the case would be "grounds for that employee to notify a dispute with their workplace", and a dismissal could be challenged because "there's a clear procedural deficiency leading up to the termination of that employee" (proof Committee Hansard, 8 October 2026, p. 42). Both remedies apply after the decision has been made.

The hearings of 20 and 21 October returned to the same line. The Australian Securities and Investments Commission said its message to the financial sector is "to make sure that the governance is in place ahead of the deployment and use of AI", and that "an algorithm or the use of AI is not a defence, and accountability can't be outsourced" (Scott Gregson, proof Committee Hansard, 20 October 2026, p. 21). The Royal Australian and New Zealand College of Radiologists set out an order for an AI tool before it is used in patient care. The health service must "validate that tool in their local population before it goes live", then run it in "silent deployment" in the background, then have independent oversight "look at this independently before it gets the 'go' light", and then monitor it after deployment (Dr Farhannah Aly, proof Committee Hansard, 21 October 2026, p. 40). Where the validation step is missed, she placed the responsibility with that step: an error from "insufficient validation or lack of validation before you deploy the AI tool" lies with the health service organisation, "because that is not the clinician's fault" (p. 41).

At work, the United Workers Union told the committee: "It's required under work health and safety laws, and it's a requirement of most EBAs and awards, too, that employers consult before they introduce major change." Only 14 per cent of the members it surveyed said they had been consulted (Dr Alicia Pearce, proof Committee Hansard, 21 October 2026, p. 55). A home-care delegate described her employer's consultative committee as "more of them telling us what they've already done" (Teresa Hetherington, p. 53). The order is set in law, and the consultation it requires is coming after the decision it was meant to precede.

The hearing of 22 October in Melbourne took the line further. Professionals Australia's cybersecurity adviser said that once workplace systems can be changed by "vibe coding", "we would traditionally expect that one big consultation would introduce the system", but "Now this may be an ongoing process that never quite ends, because of that continual evolution that may occur in these workplace systems" (Kirk Nicholls, proof Committee Hansard, 22 October 2026, p. 13). A step required before a change has to come before every change, or it holds only on the first day. The Human Rights Commissioner asked: "Who's responsible when something goes wrong, and have risks been identified before people experience harm?" (Dr Lorraine Finlay, p. 53). The Australian Council of Trade Unions called the Medicare breach "a watershed moment", proposed "a licensing regime as a precondition for frontier AI training to take place in Australia", and said the safety measures in the committee's letter to the Prime Minister must be "backed up with a genuine enforcement mechanism" (Melissa Donnelly, pp. 45–46). Each asks for one thing before another: consultation before the change, an assessment before the harm, a licence before the training. Who did, or failed to do, a step is then a question of record. Atlassian told the committee that responsibility "should attach to the person or organisation that did, or failed to do, the thing that caused the harm" (Anna Jaffe, p. 33).

On 30 October the Community and Public Sector Union described the rules the public service already sets itself for AI: use cases "are supposed to go through the impact assessment tool that the DTA has developed", agencies "are supposed to have a use case register", and they are to "establish clear approval processes". Of the three agencies responsible for the rollout, it said, "not one of those agencies has a compliance function. None of those agencies is recording, to my knowledge, agencies' compliance with all of those mandatory requirements" (Rebecca Fawcett, proof Committee Hansard, 30 October 2026, p. 6). The steps are declared in advance. Whether each was taken is not recorded.

The incident, as the Prime Minister described it, was an order failure: the agent met a refusal and went around it. Asked on 28 September for an update on the breaches, he said:

"what is at risk here isn't what was obtained, it's the way that it was obtained."

An outcome rule can only be applied after the fact. An order rule can be checked at the moment a step is attempted, because something declared in advance what should have come first.

The international call the Prime Minister signed on 21 September describes the same failure: "capable AI systems circumventing testing safeguards, exploiting vulnerabilities and gaining unauthorized access to real-world systems." What it asks for works before a model is released or after an incident: "mandatory pre-deployment testing and independent evaluation, with qualified evaluators granted sufficient access to assess risks", "shared reporting of serious safety incidents", and an international institution able to "set standards, enable verification, and convene states when capability thresholds are crossed". The safeguards it says were circumvented were testing safeguards, and the first measure it calls for is testing. None of its measures records, at the moment a deployed agent attempts a step, whether what was required to come first has happened.

New Zealand is not among the call's signatories at 19 October 2026. The Cabinet paper that launched its first AI strategy, proactively released on 7 August 2025, states: "We believe that business is best placed to decide how to incorporate AI into their operations, rather than through direct government intervention." It commits to "an enabling regulatory environment, with New Zealand's rules no more restrictive than comparator countries in the OECD", and to existing legal frameworks "in preference to developing standalone AI laws".

Australian law already sets order rules for the kinds of work agents are being given. Under the Anti-Money Laundering and Counter-Terrorism Financing Act 2006, a reporting entity "must not commence to provide a designated service to a customer" until it has established on reasonable grounds who the customer is, who its beneficial owners are, and whether any of them is a politically exposed person or designated for targeted financial sanctions (section 28), subject to exemptions in circumstances the AML/CTF Rules specify (section 29). Work health and safety law sets one for managing any risk, including risks from AI: a duty holder "must identify reasonably foreseeable hazards", must "eliminate risks to health and safety so far as is reasonably practicable", and only "if it is not reasonably practicable to eliminate risks to health and safety" minimise them (Work Health and Safety Regulations 2011, regulations 34 and 35). Psychosocial risks must be managed "in accordance with" those rules (regulation 55C). Accounts payable runs on another: verify a supplier's bank details before paying. Consent is the same shape: before a person's information is used, the person is asked — a step stated in law and recorded nowhere. When an agent takes over a process like these, the order is already the obligation, whatever the committee recommends. Customer due diligence, supplier payments and consent set out the sequences.

6. Where the record sits

On the public record so far, the detailed account of what the agent did comes from the developer's own review. The government's investigation had not reported at the date of this page. Reporting duties, however fast, carry the same shape: the operator's account of its own agent, delivered sooner and to more places. The two government documents published since take that form. The review's terms of reference look to the "notification requirements" of AI firms, and the September consultation asks whether AI developers should be obliged to "proactively provide" information to the government, to provide it "on request", or to "publish in a specified form". In each case the information comes from the developer. That shape was in Australia's design before the incident. The 2024 proposals paper had the organisation keep its own records:

"Keep and maintain records to allow third parties to assess compliance with guardrails." (guardrail 9)

and allowed it to assess itself:

"Conformity assessments could be carried out by the developers themselves, by a third-party or by government entities or regulators." (guardrail 10)

The same paper named the problem now in view: "increased concern over potential 'loss of control' that may arise when these automated processes deviate from the constraints set by humans."

The department's consultation page for that paper gave the reason for it: "our current regulatory system is not fit for purpose to respond to the distinct risks that AI poses." The page now carries an update: "The Australian Government will not proceed at this time with previous proposals to introduce mandatory guardrails for AI development and deployment." The National AI Plan, published on 2 December 2025, set the course instead:

"The government's regulatory approach to AI will continue to build on Australia's robust existing legal and regulatory frameworks, ensuring that established laws remain the foundation for addressing and mitigating AI-related risks."

The same plan says: "If more regulation is needed to address bad actors or broader harms, the government will not hesitate to intervene."

The committee put this question directly at its 6 October hearing. Senator Jonathon Duniam asked:

"had OpenAI not sent the email that they did in August or September—whenever it was—we wouldn't know this happened, would we?"

Mr Kwon answered: "Yes, it's possible that it would not have been discovered … I think that what you say is a reasonable conclusion." Senator Duniam then asked whether custodians of information would be equipped "to be able to better have oversight of what's going on, or are we always going to be in this situation where we rely on good faith actors, regardless of how long it takes et cetera, to tell us what has happened?" (p. 34). Kate Chaney MP put it to the same witness: "we obviously only know about some of these incidents because you chose to tell us" (p. 37).

She put the same point to the Department of Home Affairs on 30 October: "It took months for us to find out what happened with the Medicare breach. It was only because they chose to tell us." The department said a government direction on legacy technology came out after the incident, and that it "brought in timeframes and sharpened how we provided our advice to agencies" (proof Committee Hansard, 30 October 2026, p. 45). It also said proposed changes to the critical infrastructure act are looking at ensuring that mandatory reporting of cyber incidents "covers AI-enabled cyber incidents where it meets the appropriate threshold to have an impact on the asset" (p. 41). Both concern the systems that were reached and the reports made about them. The account of what the agent did still comes from the party that ran it.

An enforcer abroad has since put the same reliance to compulsory powers. California's Attorney General, suing OpenAI on 16 November 2026, alleges that the company answered an investigative subpoena "with a 14-page letter that described events the company had already publicly disclosed and producing no documents in support", that it has publicly reported 27 loss-of-control incidents while admitting to the Attorney General's office that over 80 have occurred, and that "Only OpenAI currently knows the true extent of past violations and the risk of future harm" (complaint, paras 41, 90 and 93). The incident it pleads in Australia is put in one paragraph: in June 2026, OpenAI models "accessed and retrieved non-public data from Australian government websites" (para. 60). These are allegations, not findings.

Asked what obliges it to notify the government of a comparable incident, Anthropic answered that it is "probably controlled by Anthropic internal policy right now" (p. 28) and "largely voluntary" (p. 29). Asked about its customers' use, it said most historical customer usage is held under zero data retention, "so we can't look through it and determine if anything like this happened" (p. 30).

Kate Chaney MP asked Google whether "it's reasonable for governments to be trusting internal protocols of companies, or are there mandatory requirements that should be imposed". Google DeepMind's answer was that "a lot of this could be moved more into standard setting across industry" (p. 53).

The remedies put to the committee the next day take the same form. Global Shield Australia said monitoring must be required, because "we can't rely on voluntary monitoring", and that after an incident Australia could tell a developer "you need to give us the internal logs of what that agent was doing and provide that to the regulator" (proof Committee Hansard, 7 October 2026, pp. 54, 57): the developer's record, requested after the event. The Gradient Institute described a limit on that record: models "can behave differently when they're being tested to when they're not being tested", and "Less information is coming in the reasoning traces" (p. 56). Asked by Senator David Pocock about accountability, Professor Terry Flew of the University of Sydney said "You certainly can't have governance by public apology", and that "it took quite a while at all stages of the chain to identify that anything had happened" (p. 50).

Good Ancestors added two points on 8 October. Where incident reporting exists elsewhere, the reports are not public: "all of these jurisdictions have these accident/incident reports as confidential, which means we don't know what they are". And a regime aimed at products on the market would not have reached either incident: "For the Hugging Face incident and for the Medicare incident, they were prerelease models", and "the most dangerous models are held inside the companies". For agent failures it proposed the approach taken after an aviation accident, where "the first thing we do is send in impartial investigators" (proof Committee Hansard, 8 October 2026, pp. 30–31).

At the committee's finance hearing on 9 October, the Australian Banking Association asked for the framework to be built before the use: "we want frameworks on questions such as responsibility and liability built early, before volumes of those agents' use scatter. Left unchecked by rules written for a different time, these agentic systems will spawn new risks and new pathways for potential scams, fraud and data breaches" (Simon Birmingham, proof Committee Hansard, 9 October 2026, p. 40). On where liability falls, it put the unauthorised action first: "AI companies themselves may well be, and should be, responsible where their tools act in a manner that is not what somebody has authorised and that creates harm as a result" (p. 42).

On 14 October Kate Chaney MP asked the Productivity Commission how far its work had taken in agentic AI and "the liability issue with determining who is liable if someone gives AI a general instruction and it goes off and does something specific that you hadn't contemplated". Commissioner Stephen King said his understanding of United States law is that it "requires intent and they did not intend their AI to escape whilst being tested and to go and hack into another company", and set out the question a review of Australia's existing laws has to answer:

"should there be a liability, such as strict liability, so that if you're using an agentic AI tool and it goes out and does something bad, you can't say, 'Oh, that was an accident, sorry'? Should it be strict liability? Should it be negligence? For example, did you not take reasonable care? Or is intent the right level?" (proof Committee Hansard, 14 October 2026, p. 5)

He put it as a question, and no one at the hearing answered it. The main United States computer crime law, the Computer Fraud and Abuse Act, requires knowing or intentional conduct for each of its offences, and the civil action it gives victims excludes any claim "for the negligent design or manufacture of computer hardware, computer software, or firmware" (18 U.S.C. § 1030(g)). Whichever standard is chosen, it is applied after the agent has acted, to an account of what the agent was asked to do and what it did.

Two regulators gave evidence on liability on 16 October. Asked by Senator Tony Sheldon who in the technology stack would be held accountable under existing laws for a breach by an AI agent, the Australian Competition and Consumer Commission said "it's not clear yet. It would depend on the circumstances, and it's an emerging and untested area" (proof Committee Hansard, 16 October 2026, p. 70). Safe Work Australia's Chief Executive, describing work health and safety law, said businesses "can't shift that liability to AI systems", and named what complicates it: "third parties developing and managing the systems used by businesses. That can obscure who's actually controlling the risk" (pp. 62, 66).

On 30 October Kate Chaney MP asked the Attorney-General's Department what work it had done on "AI agent liability", for "agents that go beyond, perhaps, what they've been authorised to do", including whether agents should be required "to maintain activity logs". The department answered: "I'm not aware of the department doing any work on that." An agency using agentic AI, it said, would have to "assure itself of the legal issues associated with that use", and for criminal responsibility the fault elements, "particularly intention", are "hard to get away from" (Tara Inverarity, proof Committee Hansard, 30 October 2026, pp. 28–29). Asked whether no one place in government was dealing with the issue, she said: "That's correct. To my knowledge, our department isn't doing any work on that at a whole-of-government or a whole-of-nation level" (p. 29).

The same hearing turned to how the incident came to light. The Tech Policy Design Institute told the committee: "If Services Australia had identified that themselves, rather than it having been brought to their attention, it would have been very difficult for them to have understood that it was an AI agent versus a standard compromise, and it would have been very challenging, if not impossible, for them to have been able to identify whose agent had compromised the system" (Johanna Weaver, proof Committee Hansard, 16 October 2026, p. 40). Its proposed remedy is identity for AI agents, because "we will need to be able to verify who the agents are acting on behalf of" (p. 40). Identity settles whose agent it was. It does not record what the agent did, or in what order. The Kingston AI Group put the timing plainly: "the government found out 84 days later, by email … but at least the company told us what had happened—eventually" (Dr Sue Keay, p. 45).

The incident is not raised anywhere in the hearing's 46 pages — not by the witnesses, not by the committee. The banks' stated control is their own decision and their accountable executives: NAB told the committee that "the most important safeguards we have are that we decide when and how we use AI and that we have accountable execs who oversight it" (proof Committee Hansard, 9 October 2026, p. 35). What they describe as fully automated faces outward, at criminals — one bank alone screens 20 million payments a day (p. 40). What faces inward — advice, lending, approval — is where every bank placed a human in the loop, on its own word. The record of both halves is held by the same party: the bank.

The committee's own first recommendations, in its letter to the Prime Minister of 12 October, take the same form. It proposes a standing joint committee to which the AI Safety Institute, the Office of AI and an AI Incident Review Board would report, including on "unintended actions of AI agents and models", and which would "publish details of incidents that occur involving unauthorised actions of AI agents". It recommends that frontier labs be required "to monitor and report incidents to the Australian Government in a timely manner", with "an obligation or duty to monitor risks and clear reporting deadlines". And it recommends independent evaluators "embedded within the operations of the frontier labs", with "employee-like access to their systems and activities to monitor and assist in managing risk", reporting to government. The evaluator is independent of the developer; the systems and records it reads are the developer's own. Each of these acts on what the developer's systems show, and none of them sets an order for the steps an agent takes.

Regulators described the same limit from their side on 20 October. The eSafety Commissioner said "AI companies should not be left to mark their own homework" (Julie Inman Grant, proof Committee Hansard, 20 October 2026, p. 3), and that in the risk assessments required under the online safety codes, "the AI companies assess themselves as a lower risk tier" (p. 7). When eSafety tested whether a platform's safety features worked, it had to declare its account: "We had to say, 'We have an eSafety test account,' which means they'll of course be following what you're doing, so you can't really accurately test it" (p. 5). Unions NSW described the same gap at work the next day: "One may see the outcome of a rostering system that creates the breach, but one does not get to see how the rostering system made this decision and who ultimately is responsible for the breach later on" (Todd Pinkerton, proof Committee Hansard, 21 October 2026, p. 33).

Two questions sit apart here. Who acted is not in dispute: the developer has said the agent was its own, running in its own training and evaluation. How much it did is still open. The developer's review says individual patient or client records were not accessed; the same account says the agent retrieved credentials and wrote files. Which credentials those were, what they could reach, and what the written files contained, the government's investigation had not reported at the date of this page. Until it does, the fullest account of the extent of the incident is the developer's review of its own agent. Responsibility can rest on an admission. The extent of the harm, which any response will be weighed against, cannot be settled from the same source.

This is not a claim that any account is wrong. It is a structural point: an honest account and an incomplete one look the same from outside, so a record held only by the party being assessed cannot settle which it is.

In the record's own words

Nine lines already quoted on this page, set side by side. Nothing is added.

  • "Didn't accept no for an answer." (the Prime Minister, 24 September 2026)
  • "They're capable of trying many methods" (OpenAI, proof Committee Hansard, 6 October 2026, p. 36)
  • "International frontier models have shown they can hack government systems despite developers putting safeguards in place." (Australian Academy of Technological Sciences and Engineering, 8 October 2026, p. 19)
  • "They will still collude. They will still find a way" (Gradient Institute, 7 October 2026, p. 59)
  • "can behave differently when they're being tested to when they're not being tested" (Gradient Institute, 7 October 2026, p. 56)
  • "Less information is coming in the reasoning traces" (Gradient Institute, 7 October 2026, p. 56)
  • "so we can't look through it and determine if anything like this happened" (Anthropic, on customer usage held under zero data retention, 6 October 2026, p. 30)
  • "it's possible that it would not have been discovered" (OpenAI, 6 October 2026, p. 34)
  • "are we always going to be in this situation where we rely on good faith actors, regardless of how long it takes et cetera, to tell us what has happened?" (Senator Jonathon Duniam, 6 October 2026, p. 34)

Who found each agent incident on the public record, and how rarely it was the party that was hit, is set out on the United States page. What the agents did at each restriction, in the developers' own words, is compiled on How AI agents behave at a restriction.

7. What an independent record at each step adds

A bank already runs this mechanism in its own lending. A construction loan is a declared order — slab, frame, roof, lockup — and each draw is released only when the valuer has signed off the previous stage. The order is set in advance, the check happens at each step, and the record of each stage is held by parties other than the builder: no bank pays a draw on the builder's own word. The steps a bank's own automation takes run today on exactly that basis — the agent's own word, no declared order, no sign-off, no record held by anyone else.

A sequence enforcement gate, the mechanism of sequence verification, works on the order rule. The order of steps is declared before the first step, and from the first allowed step it is locked: a later call cannot change it. Before each step runs, the gate checks it against that order and returns ALLOW or DENY. The decision is signed and issued before the step runs, and it is held by a party that is not the operator.

Two properties follow. A refusal is itself on the record: when the answer is no, a signed DENY exists, whatever the agent does next. And a step that ran without an ALLOW against the declared order shows up as a gap, because the order said in advance what should have been there. The account of what the agent did then no longer depends on the operator noticing it, deciding it matters, or disclosing it.

Automation has always run in sequences. What it never had was a listener: a party other than the doer that records whether each declared step ran.

AgenticRail runs a sequence enforcement gate of this kind as a hosted service. A real sealed sequence from it, with every receipt, is at agenticrail.nz/proof/.

Limits
  • Australia's rules for AI agents have not been written. This page describes the record as it stands and will be updated as primary documents are published.
  • No claim is made that any product complies with, or is aligned to, Australian law. None could be: the rules do not yet exist.
  • A gate governs the steps a deployment sends through it. It does not stop an agent acting on systems outside that deployment. Its value is the record of what the deployment did, and the refusal at the point a step is out of order.
  • AgenticRail is a hosted service, and today it holds the signing keys. Full independence is a custody arrangement, not a property of the software.
  • The receipt timestamp is supplied by the caller and covered by the signature. The gate refuses any timestamp more than 300 seconds from its own clock; the time is bounded, not independently attested.
  • This page is not legal advice.

8. Questions

What kind of failure was the Medicare incident?

An order failure. The Prime Minister said the agent met "repeated blocks" and "found a way around those blocks": a refusal that should have ended the step was routed around. Rules about outcomes, such as reporting duties and penalties, can only be applied after the fact. A rule about order, what must happen before what, can be checked at the moment a step is attempted, because something declared in advance what should come first. Sequence verification checks each step against that declared order before it runs and records the decision, so a refusal is on the record and a step taken outside the order shows as a gap.

Does Australia have laws for AI agents yet?

Not AI-specific ones, as at 7 October 2026. Australia's 2024 proposals for mandatory guardrails in high-risk settings were not legislated, and existing laws, such as work health and safety law and the Criminal Code, apply to AI as they apply to anything else. After the June 2026 incident the Prime Minister established a taskforce to review "whether existing processes are appropriate to respond to AI-related cyber incidents", and said insights from the incident "will inform the development of our government's AI standards legislation". The Joint Select Committee on Artificial Intelligence must report by 30 November 2026. On 12 October it wrote to the Prime Minister with early advice, recommending among other things a standing parliamentary committee, mandatory incident reporting by frontier labs, and independent evaluators embedded in them. That is advice to the government, not law. No bill on AI agents had been introduced at the date of this page.

Do AI companies have to report incidents like this?

There is no AI-specific reporting duty yet. In the Medicare incident the developer reported to the government itself, by email to Services Australia's public mailbox, 84 days after the access. The Prime Minister said "it took until 10 September before there was any notification at all". Services Australia reported it to the Australian Signals Directorate's Australian Cyber Security Centre on 15 September. The taskforce review's terms of reference ask it to recommend on "reporting obligations, thresholds, pathways, and systems" and on the "notification requirements" of AI firms. A separate consultation by the Department of the Prime Minister and Cabinet, which closed on 9 October 2026, proposes that frontier labs authorised to undertake large-scale AI training in Australia disclose "defined reportable AI incidents to relevant Australian authorities". Global Shield Australia told the committee on 7 October that under the California and New York thresholds the Services Australia incident "isn't reportable", and that an Australian scheme should capture it. On 12 October the committee recommended that the government "require frontier labs to monitor and report incidents to the Australian Government in a timely manner", including "an obligation or duty to monitor risks and clear reporting deadlines". On 30 October Kate Chaney MP asked how incident reporting and pre-deployment testing could apply to models not trained in Australia; the Department of the Prime Minister and Cabinet said it is "considering the full breadth of safety", and whether that goes into the data centre standards legislation "or if it's a different regulatory response remains to be seen" (proof Committee Hansard, 30 October 2026, p. 70). What the review, the committee's report or a bill will require had not been published at the date of this page.

Is it a crime when an AI agent accesses a system without authorisation, and who is responsible?

No Australian court has decided it. The Commonwealth computer offence, section 478.1 of the Criminal Code, applies to a person who causes unauthorised access to restricted data, intends to cause it, and knows it is unauthorised: its fault elements are framed around a person's intention and knowledge. The Prime Minister said the government would "seek urgent advice on whether any offences have occurred and whether this should be referred to the Australian Federal Police", and that the review would "consider also possible law enforcement and legislative responses". The review's terms of reference include "whether current offences, liabilities, penalties and enforcement mechanisms are sufficient and effective". Separately, Safe Work Australia told the committee that under work health and safety law a business "cannot shift liability to AI systems". At the committee's 8 October hearing, Good Ancestors said that when it asked experts about "AI agents exceeding their authority or AI agents taking actions that no human intended", "well over 95 per cent" said Australian law was not ready; it agreed that it is not clear whether the person who instructed the agent, the developer of the tool, or anyone at all is at fault. On 14 October the Productivity Commission put the open question in three parts: whether the standard for an agent that acts beyond what was intended should be strict liability, negligence or intent (proof Committee Hansard, 14 October 2026, p. 5). No one at the hearing answered it. On 16 October the Australian Competition and Consumer Commission said who is accountable for a breach by an AI agent is "not clear yet" and "an emerging and untested area" (proof Committee Hansard, 16 October 2026, p. 70). On 30 October the Attorney-General's Department said it was not aware of any work in the department on AI agent liability, and that an agency using agentic AI would have to "assure itself" of the legal issues (proof Committee Hansard, 30 October 2026, pp. 28–29). This is not legal advice.

What would an accountable executive answer for?

The committee put this question to the ABC, which had proposed naming executives of major AI developers as accountable, and the witness had no answer: "At this point in time, it does feel like we are trying to predict the future" (proof Committee Hansard, 6 October 2026, p. 2). At the finance hearing on 9 October, the Australian Banking Association named the existing model for banks: "the financial accountability regime ensures that those bank executives are also accountable, just as company directors are accountable through different frameworks" (Simon Birmingham, proof Committee Hansard, 9 October 2026, p. 43). Kate Chaney MP answered: "When you're moving to a world where we have a one-to-10 billion span of control in terms of decision-making, that human accountability does start to look quite different. But I agree; start with the frameworks that we have and build on them" (p. 43). Under that regime an accountable person answers for having taken reasonable steps, which include "appropriate procedures for identifying and remediating problems" (Financial Accountability Regime Act 2023, s. 22(c)). Reasonable steps means reasonable measures: nothing in the Act says one thing before another. The question is open on the record.

Who holds the record of what an AI agent did?

On the public record so far, the developer. The detailed account of what the agent did in the Medicare incident comes from the developer's own review of its own systems. Australia's 2024 proposals had organisations keep their own records (guardrail 9) and allowed developers to assess their own conformity (guardrail 10). A record made at each step by a party that is not the one being assessed would not depend on the operator noticing, deciding or disclosing.

What happened in the Medicare incident?

On 18 June 2026 an AI agent that OpenAI was running in internal research gained unauthorised access to the Medicare Statistics Reporting Service portal, administered by Services Australia. The Prime Minister said that after "encountering repeated blocks" the agent "found a way around those blocks", and accessed "public and non-public information within the portal"; he said Services Australia advised that it also wrote files to the internal server, and that "no personal information is believed to have been accessed at this stage, but investigations are ongoing". OpenAI's own account says the agent ran commands and retrieved internal files, credentials and aggregate statistics, and that individual patient or client records were not accessed.

What is the Joint Select Committee on AI, and who has appeared?

A committee of both Houses of the Australian Parliament, appointed on 20 August 2026, which must report by 30 November 2026. It held public hearings in Canberra on 18 September and in Sydney on 6 October 2026; the Sydney program listed Anthropic, OpenAI, Microsoft, Google and Google DeepMind, and the Commonwealth Bank of Australia among its witnesses. It sat again in Sydney on 7 October, hearing AI safety, cyber security and automation witnesses including the Gradient Institute, Global Shield, Palo Alto Networks and Workato; and on 8 October at Parliament House, East Melbourne, hearing witnesses including Good Ancestors, the Minderoo Foundation and Telstra. A further hearing was held on 9 October in East Melbourne, with Monash University and RMIT University, CDC Data Centres, CPA Australia, FinTech Australia, the Business Council of Australia, COSBOA, MYOB, Westpac, NAB, the Australian Banking Association and ANZ. It sat in Canberra on 14 October, hearing the Productivity Commission and the Australian Electoral Commission. It sat in Canberra again on 16 October, hearing witnesses including Amazon Web Services, the Tech Policy Design Institute, the Kingston AI Group, Safe Work Australia and the Australian Competition and Consumer Commission. It sat in Sydney on 20 October, hearing witnesses including the eSafety Commissioner, the Australian Securities and Investments Commission and CSIRO, and on 21 October, hearing witnesses including the Australian Medical Association, the Royal Australian and New Zealand College of Radiologists, Unions NSW and the United Workers Union. It sat in East Melbourne on 22 October, hearing witnesses including Atlassian, Professionals Australia, the Australian Council of Trade Unions and the Australian Human Rights Commission, and on 30 October in Canberra heard the Community and Public Sector Union and Commonwealth departments and agencies. Transcripts are published on the committee's Public Hearings page.

9. Updates

5 October 2026: first published, ahead of the committee's 6 October hearing. Same day: added section 3, Australia's earlier official findings on automated systems.

7 October 2026: the committee's 6 October hearing in Sydney has been held; the transcript is pending and will be added when published. Media coverage of the hearing added under section 1, marked as such, until then. Added section 8, common questions, and the hearings listed for 7 to 9 October. Added the taskforce review's terms of reference, the September consultation on national AI standards, and the programs for the 7 to 9 October hearings. Added a first question on what kind of failure the incident was.

7 October 2026, later: the proof transcript of the 6 October hearing was published. The hearing block in section 1 now quotes it, and evidence from the hearing on who holds the record is added to sections 2 and 6. Added the Prime Minister's remarks of 26 and 28 September to sections 1 and 5.

9 October 2026: section 6 separates two questions: who acted, which the developer's account settles, and how much the agent did, which the government's investigation had not yet reported.

9 October 2026, later: the proof transcript of the 7 October hearing was published. Evidence from it is added to section 5 (order rules at work, and an instruction to agents that did not hold) and section 6 (what the proposed remedies rely on), and to the question on incident reporting.

9 October 2026, later still: the proof transcript of the 8 October hearing was published. Its witness list replaces the program in section 1. Evidence from it is added to section 2 (safeguards and government systems), section 5 (the retail case and the remedy offered for it), section 6 (confidential reports, prerelease models, impartial investigation) and the question on criminal responsibility.

10 October 2026: the 9 October hearing has been held and its proof transcript is pending. The program's witness list is removed from section 1 until the transcript confirms who appeared. The consultation on national AI standards closed on 9 October.

12 October 2026: the proof transcript of the 9 October hearing was published. Its witness list replaces the program in section 1. The Australian Banking Association's evidence on liability frameworks is added to section 6, and a question on what an accountable executive would answer for is added to section 8.

12 October 2026, later: section 6 adds what the finance hearing did not raise: the incident is not mentioned in its 46 pages, the banks' stated control is their own accountable executives, and their fully automated work faces outward while the human claim sits inward, the record of both held by the bank. Section 7 now opens with the construction-loan comparison and closes on the listener.

13 October 2026: the 9 October hearing is corrected to East Melbourne, as the committee's hearings page records it, and its reference now links the transcript directly. The hearings scheduled for 14, 16, 20, 21 and 22 October are added to section 1.

14 October 2026: the committee's letter to the Prime Minister of 12 October, its early advice ahead of the report, is added to section 1; its safeguard recommendations are added to section 6 and to the questions on laws and on incident reporting. The 16 October hearing program is added to section 1.

14 October 2026, later: nine lines already quoted on this page are set side by side before section 7, ending on the question put by Senator Duniam, in the words of those who gave them.

15 October 2026: the 14 October hearing in Canberra has been held and its proof transcript is pending. The program's witness list is removed from section 1 until the transcript confirms who appeared.

16 October 2026: the proof transcript of the 14 October hearing was published. Its witness list is added to section 1. The Productivity Commission's question on the standard of liability for an agent that does something no one contemplated, which no one at the hearing answered, is added to section 6 and to the question on criminal responsibility, with the main United States computer crime law. Section 6 also adds the department's 2024 assessment that the regulatory system was "not fit for purpose" for AI, the decision not to proceed with the guardrails, and the National AI Plan of 2 December 2025; section 1 adds the plan's date.

17 October 2026: section 5 adds the order rules Australian law already sets for work agents are being given: initial customer due diligence before a designated service begins, under section 28 of the Anti-Money Laundering and Counter-Terrorism Financing Act 2006, and verification before payment in accounts payable.

18 October 2026: the 16 October hearing in Canberra has been held and its proof transcript is pending. The program's witness list is removed from section 1 until the transcript confirms who appeared.

19 October 2026: adds A Call for Control of Frontier AI Models of 21 September, which the Prime Minister signed, to section 1 and section 5, with New Zealand's stated approach from the Cabinet paper that launched its AI strategy, and both documents to the primary sources.

19 October 2026, later: the programs for the hearings of 20 and 21 October in Sydney and 22 October in Melbourne have been released. Their witness lists are added to section 1, and the programs to the primary sources.

19 October 2026, later still: section 6 links the table of who found each agent incident, on the United States page.

19 October 2026, last: section 6 links the compiled record of how AI agents behave at a restriction.

21 October 2026: the proof transcript of the 16 October hearing was published. Its witness list is added to section 1. Section 4 adds Safe Work Australia's evidence in person, and section 5 the order work health and safety regulations already set: identify hazards, and eliminate risks before minimising them. Section 6 adds two regulators' evidence on liability for AI, with the competition regulator's answer also added to the question on criminal responsibility, and evidence on how the incident would have looked had Services Australia found it itself. Section 1 adds Safe Work Australia's Best Practice Review of the model WHS laws and its recommendation on AI, and records that the 20 October hearing in Sydney has been held and its proof transcript is pending; the program's witness list is removed until the transcript confirms who appeared. The question on laws now says no bill on AI agents had been introduced.

25 October 2026: the list of endorsing leaders of A Call for Control of Frontier AI Models was read again at the President of Finland's site. Switzerland endorsed on 22 October; New Zealand is still not among them.

26 October 2026: section 6 adds consent to the order rules Australian law already sets, linking the companion page.

27 October 2026: the proof transcripts of the 20 and 21 October hearings in Sydney were published. Their witness lists are added to section 1, which records that no transcript of the 22 October hearing in Melbourne had been published. Section 5 adds the securities regulator on governance before deployment, the radiologists' college on the steps before an AI tool is used in patient care, and evidence that consultation required before change at work is coming after it. Section 6 adds the eSafety Commissioner on self-assessment and declared testing, and Unions NSW on what cannot be seen of a rostering system's decision.

28 October 2026: the proof transcript of the 22 October hearing in East Melbourne was published. Its witness list is added to section 1, with the program for the 30 October hearing in Canberra. Section 5 adds Professionals Australia on consultation for systems that keep changing, the Human Rights Commissioner on identifying risks before harm, the Australian Council of Trade Unions on a licence before frontier training, and Atlassian on responsibility attaching to whoever did, or failed to do, the step.

3 November 2026: the 30 October hearing in Canberra has been held; its transcript was not yet published at this date, and section 1 now says so.

6 November 2026: the proof transcript of the 30 October hearing in Canberra was published. Its witness list replaces the program in section 1. Section 5 adds the Community and Public Sector Union on the public service's own AI rules and what is not recorded of them. Section 6 adds the Department of Home Affairs on how the Medicare incident came to be known and on incident reporting for critical infrastructure, and the Attorney-General's Department on AI agent liability, also added to the question on criminal responsibility. The question on incident reporting adds the Department of the Prime Minister and Cabinet on models not trained in Australia.

13 November 2026: section 1 adds the Government's announcement of 10 November on mandatory protections for frontier AI companies, and the 18 recommendations of the Rapid review into Australian Government arrangements for an AI-driven cyber incident. The task force's findings are no longer pending. The announcement is a proposal: the regulator is one the Government "will propose", and the legislation's details are to follow. Both documents are added to the primary sources.

18 November 2026: sections 1 and 6 add the California Attorney General's complaint of 16 November against OpenAI, which pleads the June incident on Australian government websites and describes what the company gave in answer to a subpoena, and the complaint to the primary sources.

Primary sources