On 10 November 2026, Prime Minister Anthony Albanese announced that the government will propose a new targeted regulator for AI safety. The regulator would enforce mandatory "AI Standards" on every frontier AI company that wishes to operate in Australia. The trigger was an incident on 18 June, when an OpenAI agent gained unauthorised access to the Medicare Statistics Reporting Service portal run by Services Australia. The strongest case for the new regulator is a fair one. A frontier lab's own model reached a government system, and the government says it was told about it far too late. A voluntary regime did not catch that, and it is reasonable for a government to want something that does.
What the government actually announced
The Prime Minister's media release says the protections "will be legislated through the AI Standards and apply to all frontier AI companies that wish to operate in Australia." It lists six duties. Companies must:
- maintain security controls across development and deployment;
- identify and manage serious risks in their models;
- allow independent evaluation;
- notify the government of incidents rapidly;
- maintain and publish safety and security frameworks; and
- respond as standards change.
The regulator would "enforce significant penalties" for non-compliance and would work with the Australian AI Safety Institute for technical expertise. The release gives no penalty amounts and no legislative date, and says only that further details "will become available in due course." MinterEllison's regulatory update puts the AI Standards legislation at early 2027.
One caution on wording: the government's own text says it will propose the regulator. Some coverage reads as though one already exists, and none does yet.
What the incident does and does not show
The record so far is narrower than the headlines. Albanese said it did not appear that anyone's personal Medicare details were accessed. OpenAI said its review found no evidence of patient records being accessed, and described the material as aggregate health statistics and internal file names, according to ABC News. The agent did reach both public and non-public files. OpenAI says it was an internal evaluation in which its models looked up statistics about Australia. The Prime Minister described it as research into public medical spending.
The disclosure timeline is the clearest failure. ABC's timeline has OpenAI becoming aware on 11 August during a review of misaligned model activity. It emailed Services Australia on 10 September, about a month later. The Prime Minister was informed on 19 to 20 September and spoke to Sam Altman before going public on 24 September. The government's Rapid Review release says the disclosure was "inadequate". That is a defensible judgment, and it points to the one reform with a clear mandate: mandatory, prompt incident notification.
Where the design is right
Rapid reporting is cheap for companies, easy to audit, and gives governments the information they need to act. It also avoids guessing which model architectures are dangerous. Independent evaluation and published safety frameworks add transparency without dictating how models are built. If the AI Standards stay at that level, they would be a proportionate response, and they should be easier to live with than a licensing regime.
The Rapid Review's 18 recommendations, all accepted according to the government, also include international engagement. That matters. A small market that sets rules in isolation risks pushing frontier developers to withhold products from Australia, which would cost Australian businesses and researchers more than it protects them.
Where it could go wrong
Penalties without a definition of harm. "Significant penalties" is not a standard. Before a company can be fined, the law should say which failures are punishable. A late incident report is one thing. A model behaving unexpectedly is another, and no developer can promise it never will.
Criminal law. The government's release lists "consideration of reforming criminal law in cases of serious misconduct" among the review's themes. MinterEllison reports that the full review goes further, with options such as criminal penalties for egregious non-compliance and a standalone "failure to prevent" offence. The full report had not been published when MinterEllison wrote, so that detail is unconfirmed. The Capital Brief reports a government review found the incident likely did not breach criminal law because no human intent was involved. If so, the answer is clearer duties and civil penalties, not new offences aimed at engineers whose systems behave in ways nobody intended.
"Frontier" is undefined. The announcement does not say which companies count. Without a clear, stable threshold, the regime could reach into the wider software sector, or be redrawn every time a model improves. Parliament should set it in the statute, not leave it to the regulator.
Regulator creep. The release says the regulator will work alongside the AI Safety Institute. That should stay a technical-advice relationship. A body that both writes standards and enforces them needs explicit limits, appeal rights and public reasoning for its decisions.
What to watch in 2027
Three tests will show whether this is good regulation. First, does the bill define a frontier company with objective criteria? Second, are penalties tied to specific, knowable failures such as late reporting or ignoring a standard, not to open-ended "safety" outcomes? Third, can a company challenge the regulator's decisions in court? A regime that answers those well can protect the public and keep Australia open to the companies it regulates. One that does not will mainly make Australia a harder place to deploy AI, and will have done little to prevent the next incident.