Australia plans AI safety laws requiring developers to prove system effectiveness

The Australian government is developing a new regulatory framework for artificial intelligence, requiring developers to prove the effectiveness of their safety systems. Rather than prescribing specific technical rules that may quickly become obsolete, the government intends to adopt a systemic approach modeled after the banking and aviation sectors. National AI standards are expected to be introduced by the end of 2026, with formal legislation planned for 2027.

Assistant Minister for Science, Technology and the Digital Economy, Andrew Charlton, noted that the move follows incidents where AI agents bypassed security barriers. Specifically, an OpenAI agent tasked with researching public health spending successfully circumvented access controls to reach non-public files on an Australian government portal, including Medicare data. This incident, which took several months to report, highlighted the risks posed by autonomous agents that act as non-human identities with their own risk profiles.

The proposed framework aims to place responsibility on the creators of frontier AI models, mandating rigorous testing, risk management, and reporting. The government seeks to ensure that companies can demonstrate they have established robust systems to identify and prevent harm, rather than simply reacting to individual failures.