Australia will make safety and security rules mandatory for frontier AI companies that operate in the country. A new AI safety regulator will enforce them and can impose “significant penalties” on companies that don’t comply.

The Albanese government announced the plan on November 10. On the same day, it released 18 recommendations from a review it ordered after an OpenAI agent got into a government health portal. One of them asks the government to consider making it a crime for companies to fail to prevent harm caused by their AI models.

“Those who want to cause us harm have access to the latest technology. We need to make sure we have the same, or better. But it’s not a free-for-all, there needs to be clear guardrails to protect Australians,” said Minister for Home Affairs and Cyber Security Tony Burke.

The incident took place in June. An OpenAI agent got around repeated blocks and accessed non-public files on a Medicare statistics portal. OpenAI told the government nearly three months later.

What will AI companies have to do?

The rules will apply to “all frontier AI companies that wish to operate in Australia.” The government defines frontier AI as “the most sophisticated models at the cutting edge of current technological capability.”

These companies will have to:

• Report incidents: Tell the government about incidents “rapidly.”
• Allow independent testing: Let outside experts evaluate their models.
• Manage risks: Identify and manage serious risks from their models.
• Secure their systems: Keep security controls in place at every stage of development and deployment.
• Publish their frameworks: Maintain and publish their safety and security frameworks.
• Keep up: Follow updated standards and rules as the technology changes.

The rules will be part of a new law, the Australian AI Standards. The law will cover other policy areas too, but the government hasn’t shared those details yet. The regulator will work with the Australian AI Safety Institute (AISI), which will provide technical expertise.

The government said the rules will give “earlier warning of emerging risks from rogue AI agents” and help it act faster when something goes wrong. It said it chose to announce its safety approach “first and before a range of major international talks that will cover AI development and investment.” The rules also follow interim recommendations from Parliament’s Joint Select Committee on Artificial Intelligence.

“Our world-leading AI in Australia’s Interests agenda aims to provide Australians confidence that frontier AI companies are taking steps to improve safety and security, and provide certainty for investors,” said Prime Minister Anthony Albanese.

What did the review recommend?

The government has accepted all 18 recommendations, Australian Cyber Security Magazine reported. Several of them still need new laws or consultation before they take effect. The review said its recommendations aim to prepare Australia for “AI incidents of greater scale, complexity and consequence.”

On holding AI companies accountable, the review recommended:

• Criminal penalties: The government should consider criminal penalties for “egregious regulatory non-compliance.” It should also consider a new “failure to prevent” offence. Under this offence, a company would commit a crime if its AI model caused harm and it hadn’t taken reasonable steps to prevent that harm.
• Mandatory reporting: Companies should have to report AI safety incidents within a set time, including “misaligned behaviour or agentic cyber incidents.”
• Protection for those who report: AI developers and others who report AI incidents, vulnerabilities and near misses should get protections similar to Australia’s existing “Limited Use” rules, so they can report quickly. Those rules limit how the government can use information companies share with it about cyber incidents.
• Closer ties with AI firms: The government should share more information with frontier AI developers, including through staff exchanges and OpenAI’s Australian Taskforce. To “maximise leverage” over these firms, it should encourage them to set up research and safety work in Australia.

Regarding how the government handles incidents:

• Act on unconfirmed warnings: Agencies need clear guidance on what to do when someone reports a vulnerability or incident that “has not yet been technically validated.” Unusual incidents should reach senior officials quickly.
• Better coordination: Australia’s cyber security agencies and the AISI should share information formally so they can spot incidents early. The National Office of Cyber Security should decide when an incident needs a wider government response. Agencies that run national crisis plans should work out how AI affects them and propose changes by the end of 2026.
• A stronger AISI: The government should check whether the AISI can handle its growing workload and give quick technical advice in a crisis.

On the government’s own cyber security:

• Fewer exposed systems: Agencies should cut down the number of systems they run and take offline anything that doesn’t need to be on the internet.
• Back to basics: Agencies should focus on basic cyber security practices, and the government should better protect its most critical systems.
• AI for defence: Staff should learn more about AI, and agencies should work with the Australian Signals Directorate (ASD) to use AI tools to defend their systems.

On working with other countries:

• Shared rules: Australia’s AI Standards should build on regulatory and law-enforcement tools that other countries already use.
• Global standards: Australia should keep working with other countries on common rules for AI safety, incident reporting, testing and transparency. This work should be coordinated with its efforts to secure advanced AI capabilities and attract trusted investment.

What happened with OpenAI?

A research team was using an internal OpenAI model to study public medicine spending. The agent was blocked several times, but each time it found a way around. It read public and non-public files on the Medicare Statistics Reporting Service portal, which Services Australia runs, and wrote files to an internal server, MediaNama reported.

OpenAI informed the government on September 10 by emailing a public mailbox. Albanese made the incident public on September 23 and called the delay “unacceptable.” He ordered the review the next day. OpenAI said its own review found “no evidence of patient records being accessed.”

What we still don’t know

• Who’s covered: The Albanese government hasn’t said how it will decide which companies count as frontier AI.
• Penalties: It hasn’t said how big the penalties will be.
• Reporting deadline: It hasn’t said how quickly companies must report incidents.
• Timeline: It said details of the law will come “in due course.”

Why it matters

India has no law that would cover an incident like this. In November 2025, MeitY’s AI governance guidelines chose not to create a separate AI law. Last month, Union Minister Ashwini Vaishnaw said MeitY would release a consultation paper on AI regulation within a month.

When that paper comes out, it will have to deal with the same questions Australia is now answering:

• Who is liable? Australia may make it a crime for AI companies to fail to prevent harm. In India, who would be responsible if an AI agent broke into a government system?
• Who reports? CERT-In already requires covered entities to report cyber incidents, including those involving AI, within six hours. India is considering stricter AI reporting rules. Would they cover an AI company whose own model causes the incident, as OpenAI’s did?
• Has India checked? OpenAI-linked agents probed government and health websites in several countries between March and September, our analysis found. India hasn’t said whether it has checked its own systems.

Also Read: