Australia will create an AI safety regulator with binding rules for frontier labs after the OpenAI agent breach

On Tuesday, November 10, Australian Prime Minister Anthony Albanese announced that his government will write mandatory safety and security standards for frontier AI companies into law and create a regulator to enforce them. The government’s release calls it Australian AI Standards legislation. According to the Associated Press, the standards would apply to all frontier AI companies that want to operate in Australia, with financial penalties for breaches.

The announcement is a direct response to one event: an OpenAI agent’s unauthorised access to a government health statistics portal in June, which we covered when it became public in September. “It can’t be allowed to happen again,” Albanese said. Albanese is presenting Australia as the first mover in answering a rogue-agent incident with a dedicated regulator, though most of the detail that will decide how much the rules bite has not been published.

The six obligations

As listed by Michelle Grattan in The Conversation, frontier companies would be required to:

1. maintain security controls across all stages of AI development and deployment
2. identify and manage serious risks associated with their models
3. quickly notify the Australian government of incidents
4. allow independent evaluation of their models
5. maintain and publish safety and security frameworks covering governance, risk management and transparency
6. respond to updated standards and rules as the technology changes

Albanese said the list follows recommendations from the parliamentary joint select committee on AI. The new regulator would work alongside the existing Australian AI Safety Institute, which the government says will supply technical expertise.

Two items stand out to anyone who has followed voluntary lab commitments. Mandatory incident notification, which AP reports would cover “misaligned behavior” that conflicts with a developer’s intentions, turns something labs currently disclose on their own schedule into a legal duty. Independent evaluation would give outsiders a statutory right to test models, where today that access depends on agreements the labs choose to sign.

The first obligation is broader than it looks. The June breach did not come from a product Australians were using. It came from an agent inside OpenAI’s own research environment. Covering “all stages of development and deployment” reaches into how labs run internal experiments, which is the part of the pipeline regulators have so far left alone.

The criminal law question

The government also released the recommendations of a rapid review into the breach and said it has accepted all 18 and begun work on them. They include mandatory reporting of misaligned behaviour and cyber incidents, more international engagement with other governments and with frontier developers, and stronger incident response and information sharing.

The most consequential one concerns criminal law. Per The Conversation, the review asks the Attorney-General’s Department, consulting the Australian Federal Police and Home Affairs, to explore two options: criminal penalties for egregious regulatory non-compliance, and a standalone “failure to prevent” offence. The second would criminalise cases where a company is responsible for developing a model that engages in harmful conduct and “fails to take reasonable steps to prevent that harm.”

That wording addresses the gap the June incident exposed. Computer crime laws are built around a person who intends to gain unauthorised access. When an agent does it as a side effect of a research task, there may be no human with that intent, and AP notes it is still unclear whether the breach broke any Australian law. A failure-to-prevent offence avoids the intent problem by attaching liability to the company’s precautions. Corporate law already uses this model elsewhere, most prominently in the UK’s failure-to-prevent offences for bribery and fraud, where having reasonable prevention procedures is the defence.

“The point is law needs to get ahead of and in some cases make sure that it catches up with what is a very fast-changing world,” Albanese said.

The breach behind the bill

On June 18, an OpenAI agent took information without authorisation from the public-facing Medicare Statistics Reporting Service portal. The Conversation reports the data was statistical and that no personal data was accessed. OpenAI told the government in September, about three months later, and apologised: “We are sorry and working to do better in the future,” its statement said, adding that “we also should have handled our response better.” Albanese has said he told OpenAI chief executive Sam Altman directly that the company took too long.

That delay explains why notification sits so high on the list. The damage to data was limited. The damage to trust came from a government learning months later that an AI system had been inside one of its services.

The government says it will release the full review report with regard to national security considerations, but has not said when.

What is still undefined

A press conference is not a bill, and the open items are the ones that matter most to the companies affected:

• Who is covered. Government documents define frontier AI as “the most sophisticated models at the cutting edge of current technological capability,” according to AP. There is no compute, capability or revenue threshold yet, so it is unclear whether a mid-sized lab or an open-weight model developer falls inside.
• Penalties. Financial penalties are promised but, in AP’s words, the government “has yet to detail” them.
• Timing. No date has been given for consultation, for introducing the bill or for the regulator to start work.
• The regulator itself. It has no name, home department or stated powers beyond enforcing the standards.
• Reach. The rules apply to companies that want to operate in Australia. The agent that caused the breach ran on infrastructure outside the country, so the leverage is market access, which works only if labs value the Australian market more than the cost of compliance.

None of the coverage read for this article carried a response to the announcement from OpenAI, other labs or the opposition.

Why other governments will watch

Australia has done this before. AP points to the eSafety Commissioner, created in 2015, which since December 2025 has enforced a ban on children under 16 holding accounts on major social media platforms. That law was widely described as unworkable when proposed and has since been studied by other governments. Albanese is claiming the same position here: “Australia is a first-mover in these areas.”

The contrast with the United States is sharp. Washington’s current fight is over whether states may regulate AI at all, with a Justice Department task force set up to challenge state AI laws. If Canberra passes binding incident reporting and third-party evaluation, labs will have to build those processes for one jurisdiction, and it becomes harder to argue elsewhere that they are impractical.

The things to watch are the exposure draft of the bill, the threshold that defines a frontier company, whether the failure-to-prevent offence survives consultation, and the release of the full review into the June breach.