AI-generated imageAustralia’s government said on Tuesday that it would legislate mandatory safety and security obligations for frontier AI companies that want to operate in the country, and would propose a new regulator to enforce them. Of the six obligations listed in a joint statement from Prime Minister Anthony Albanese and three ministers, one stands apart from the rest. Companies would have to “allow independent evaluation of their models.”
The other five are to maintain security controls through development and deployment, identify and manage serious risks, notify the government of incidents “rapidly,” maintain and publish safety and security frameworks, and keep up with standards as the technology changes. The regulator “would be able to enforce significant penalties” on companies that fail to comply, the statement said, and would work with the Australian AI Safety Institute.
It is a turn from the course Canberra chose a year ago. In September 2024 the government floated ten mandatory guardrails for high-risk AI alongside a voluntary standard. In its National AI Plan in December 2025 it set the mandatory proposal aside. The plan, GovInfoSecurity reported, asked companies to consider safety measures and said existing laws on privacy, copyright and health care would be adapted as needed. The AI Safety Institute, announced the month before, had no enforcement power. Industry groups welcomed the clarity; the Greens senator David Shoebridge reportedly called the plan “another toothless tiger.”
The clearest account of what changed since came in a speech. On October 8, Andrew Charlton, the assistant minister for science, technology and the digital economy, told a trust-and-safety audience in Sydney that “no company should release a frontier AI model that is not safe,” according to Region Canberra. He said the rules would borrow from the way banking and aviation are supervised. “Systems regulation places the onus on companies to build and run a rigorous process,” he said. The test, in that model, is whether a company has a serious system for catching risks, not just whether something went wrong.
Australia would not be the first to make frontier developers publish safety frameworks or report incidents. California’s SB 53, signed in September 2025, requires large frontier developers to post a framework and sets up a channel to report critical safety incidents to the state’s Office of Emergency Services, with civil penalties enforced by the attorney general, according to the governor’s office. That summary lists no duty to admit outside evaluators. Set beside it, the Australian item stands out. A published framework is a company’s own description of its controls; an incident report describes something that has already happened. An independent evaluation puts someone other than the company in the position of examining the model itself.
It is also the item the statement says least about. It does not say who would be allowed to evaluate—the Safety Institute, accredited private testers or the regulator itself—or when: before release, after, or both. It does not say what access evaluators would get, such as model weights, an interface or unreleased versions; who would set the scope of a test; whether results would be published or stay between evaluator and government; or what would follow a failed evaluation. “Rapidly” carries no deadline in the statement, and “incident” has no definition. Details of the legislation “will become available in due course,” the statement said, as the government “completes its detailed work and consultation.” It gave no bill text, introduction date or start date.
Those open questions decide what the duty is worth. If the company being tested chooses the evaluator and sets the scope, the obligation would leave the company deciding who looks and what gets reported. If the state chooses the evaluator, sets the scope and receives the results, Australia would have a check that does not depend on the company’s say-so. Tuesday’s six-word line can be read either way.
The statement also warns of “rogue AI agents.” Some of what that phrase points to would show up only after release, in what agents do once connected to tools and accounts. An evaluation, however independent, examines a model at one moment; an incident report describes what an agent did. How the consultation defines an incident may matter as much as who runs the tests.
Read from the side that gets tested, what is new in Tuesday’s statement is its premise: that the people who build a model should not be the only ones who get to look. The law has yet to say who else will.
Sources
Links in the text lead to the original material the journalist used.
- joint statement pm.gov.au
- floated ten mandatory guardrails gradientinstitute.org
- GovInfoSecurity reported govinfosecurity.com
- announced uts.edu.au
- Region Canberra region.com.au
- signed in September 2025 gov.ca.gov
Revision history
No revisions.
Discussion
Questions, counter-arguments and sourced corrections are welcome. Comment policy
Loading comments…