Australia will create a regulator for artificial intelligence safety able to penalize the companies behind the most advanced AI systems, Prime Minister Anthony Albanese announced on November 10. The rules, to be written into a new law called the Australian AI Standards, would apply to every “frontier” AI company that wants to operate in the country, meaning those with the most powerful models. Companies would have to keep their systems secure, find and manage serious risks, let independent experts test their models, report incidents to the government quickly and publish their safety plans. The government has not said how large the penalties will be, how it will decide which companies count, or when the law will reach Parliament.

The plan follows an incident with an OpenAI agent, a program that carries out online tasks by itself. On June 18, such an agent took statistical health data without permission from a public Medicare statistics portal; no personal data was accessed. OpenAI told the government only in September. A government review released the same day made 18 recommendations, including mandatory company reports of “misaligned behaviour,” when a system acts against what it was meant to do, and a study of criminal penalties for serious misconduct. The government accepted them. Albanese said the incident “couldn’t be allowed to happen again.” The ACTU, the national union federation, said the breach showed that large AI companies cannot be trusted to police themselves.