Australia’s government has announced plans for mandatory safety rules requiring developers of the most advanced AI models operating in the country to submit to independent testing and audits. The proposed legislation would also require ongoing safety checks, public safety frameworks and immediate reporting of security incidents to the government.
The measures would target frontier models with the highest risk profiles, rather than impose the same requirements on every AI system. Announced by Prime Minister Anthony Albanese’s government under its “AI in Australia’s Interests” agenda, the proposal would introduce mandatory safety obligations for frontier AI companies for the first time in Australia.
A new oversight body would enforce compliance alongside the Australian AI Safety Institute. Companies that breach the requirements would face substantial penalties, although the announcement leaves further details of the regulatory framework to follow after consultations conclude.
Independent access is a central element of the proposal. Frontier developers would have to make their models available for external assessment, allowing safety scrutiny beyond the evaluations companies conduct themselves. Publishing safety frameworks would also expose their approaches to identifying and managing risks to wider examination.
The requirement for continuing checks would extend oversight beyond a single assessment before a model becomes available. That distinction matters technically because an AI system’s behavior depends not only on the underlying model, but also on the tools, permissions and operating environment connected to it. Tests conducted under one set of conditions cannot establish how a system will behave in every later deployment.
Immediate incident reporting is intended in part to help authorities identify early warning signs involving AI agents that behave outside their intended limits. Such systems can combine model-generated decisions with actions through software tools. Their safety therefore involves both the content they generate and the actions they can carry out, making access controls and monitoring important parts of evaluation.
The government’s announcement was made by Albanese, Deputy Prime Minister Richard Marles and ministers Tony Burke and Tim Ayres. They framed domestic safeguards as a step to take before international negotiations, with the stated objectives of protecting Australians and national interests while establishing conditions for investment and beneficial uses of AI, including in healthcare and the wider economy.
The proposal combines several distinct forms of oversight: independent assessment, continuing monitoring, disclosure of safety practices and government notification when incidents occur. These address different stages of a model’s development and use, rather than treating safety approval as a one-off exercise.
Important implementation questions remain unresolved in the available announcement, including the precise threshold for determining which models fall within scope, how independent assessments would be conducted and the size of penalties. Further details are expected once the consultation process is complete.
Sources: The Decoder, Light Reading