Australia targets frontier AI with six mandatory safety guardrails

New rules will target frontier AI companies

Prime Minister Anthony Albanese has announced that the Australian government will establish a new AI safety regulator to enforce mandatory regulations. The regulator will have the power to impose significant penalties.

Under the announced guardrails, frontier AI companies, described in the source as companies developing the most powerful AI models at the cutting edge of AI, will have to implement security controls and identify and manage serious risks. They will also need to allow independent evaluation of their models and rapidly notify the government of serious incidents.

The remaining requirements are to publish safety and security frameworks to improve governance and transparency, and to respond to updated standards and rules as technologies change.

Policy follows an AI agent incident

The guardrails are one of 18 recommendations from a rapid review launched by the government in September. The review followed an incident in which an AI agent deployed by OpenAI to collect data accessed non-public Medicare stats.

The review also recommended criminal law reform, international cooperation and stronger mechanisms for government engagement with frontier AI developers. Nearly half of its recommendations concern increasing Australia’s preparedness for AI-related cyber incidents.

The announcement forms part of a broader technology law and policy agenda. That agenda includes a prohibition on unfair trading practices focused on online markets, a digital duty of care bill addressing harms from online platforms and AI chatbots, draft privacy law reform and an inquiry into Australia’s sovereign AI capability.

Critics warn that everyday AI remains outside the focus

Henry Fraser, a law lecturer at Queensland University of Technology, says the government is taking too narrow a view of AI safety by concentrating on frontier models. AI and algorithmic systems are increasingly used in healthcare, insurance, housing, finance, employment, education and social security, he writes. These systems are generally smaller and more specialised, such as recruitment bots, budgeting apps and educational grading algorithms.

Fraser points to examples including Robodebt, sexist AI-enabled recruitment tools, AI-enabled stadium security systems, racist policing algorithms and AI meal planners that recommend poisonous recipes. He also argues that risks can arise from interactions between software, organisations and real-world settings, as the Medicare incident showed.

Australia’s earlier 2024 proposal included ten guardrails for non-frontier AI used in high-risk settings, as well as frontier systems. Those measures included transparency, accountability, testing, data governance, record keeping, risk management, human oversight and ways to contest AI decisions. Fraser says most of those measures are absent from the new list.

Implementation will determine the policy’s reach

One potentially significant requirement is for companies to respond to new standards and rules. Fraser says that, depending on implementation, Australia’s Voluntary AI Safety Standard could offer more detailed guidance on fairness, transparency and accountability. A proposal for criminal liability for negligent AI developers could also encourage better record keeping and more active risk management.

However, Fraser says the scope of the new regulation will remain unclear until the government provides more detail on the guardrails and the rapid review recommendations. In his assessment, regulating frontier risks such as “rogue” AI agents is “good politics, and sound policy,” but does not by itself protect Australians from serious harms caused by systems already used in everyday services.