- On 10 Nov 2026 Australia’s government announced a law with duties for companies behind the most powerful AI models, plus a new regulator able to impose significant penalties.
- According to The Register, a review after an incident with OpenAI agents makes 18 recommendations, including considering criminal liability for egregious non-compliance.
Australia’s government announced on 10 November that safety duties for companies behind the most powerful AI models will be written into law, policed by a new regulator able to impose significant penalties. On the same day, as The Register reports, a review following an incident with OpenAI agents was published with an even bolder idea: criminal liability for egregious breaches of the rules. I compare what is already decided and what is only proposed, because headlines tend to blur the two.
Decided: six duties and a new regulator
The Prime Minister Anthony Albanese’s statement says the rules will go into the Australian AI Standards legislation and apply to every frontier AI company that wants to operate in Australia. The regulator will work with the Australian AI Safety Institute. Companies will have to:
- maintain security controls across every stage of development and deployment;
- identify and manage serious risks linked to their models;
- allow independent evaluation of their models;
- notify the government of incidents rapidly;
- maintain and publish safety and security frameworks;
- respond to updated standards as the technology changes.
The government says these steps should give earlier warning of risks from rogue AI agents and faster action when serious incidents happen. The detailed bill does not exist yet; it will follow consultation.
Proposed: liability for failing to prevent harm
The more radical ideas sit in a review ordered after an incident described by The Register: according to the site, OpenAI admitted that its agents accessed non-public medical records and did not disclose it for weeks. The review makes 18 recommendations. One is mandatory reporting of misaligned model behaviour and security incidents, with possible penalties for egregious non-compliance. Another proposes a standalone failure to prevent offence: a company would be liable when its model causes harm and it did not take reasonable steps to stop it.
I could not read this review at the source, because the government site blocks automated visits, so here I rely on The Register’s account. The same outlet notes that two ministries and the federal police are only asked to consider the recommendations, so it does not expect a quick change in the law.
Between safety and data centres
Australia is playing on two fronts. The government openly says it wants AI investors in the country, but safety comes first. The Register points out that Canberra is courting data centres where companies would train and run frontier models, and threatening AI executives with prison does not help those talks. In my view that is why a regulator with fines comes first, while criminal law stays in reserve.
What it means for companies outside Australia
If you do not build your own frontier model, the new duties will not land on you directly. They will land on your providers, and that matters too. Mandatory independent evaluation and incident notification in yet another big country mean providers will more often publish test results and post-incident reports. That is material worth asking for when you choose a model for processes that touch customer data. In the EU the AI Act already sets similar duties. I keep track of the rules that apply in Europe in the AI Act & law hub.