AI safety standards are coming. What changes for buyers?

If your business buys AI and does not build frontier models, nothing published so far puts a new legal duty on you under the national AI standards. They are aimed at large data centres and the labs that train the biggest models. What applies to you is the law you already have: the Privacy Act, Australian Consumer Law, workplace safety law and anti-discrimination law.

What does change is what you can reasonably expect from your vendors, and what your own customers and board will start expecting from you. The government's model asks the labs to prove they have a working process for finding and managing risk. Ask the same of the vendor selling you an AI tool, then do the same in your own business.

What was actually announced?

On 8 October 2026, Assistant Minister Andrew Charlton spoke at the Sydney Trust and Safety Festival about how the government plans to regulate frontier AI. Three things in that speech are on the record:

1. The government has committed to legislating National AI Standards for frontier AI. In his words, the government has "committed to legislating National AI Standards for frontier AI".
2. It will be systems regulation. The government will set the standard a company's risk process has to meet, then hold the company to whether that process works. It will not write a list of banned behaviours.
3. The burden goes on whoever creates the risk. His summary: "Australia's answer is to put the burden where the risk is created, to build accountability into the whole system, and to build our own capability alongside it."

He named the model's sources directly: workplace health and safety, security of critical infrastructure, prudential supervision of banks, and parts of aviation safety. And he described the test the regulator will apply:

"The goal is to ensure companies are meeting expectations of safety, without trying to specify every hazard."

The question, he said, is whether there is "a serious system in place to detect risks and prevent incidents occurring", and not only whether something went wrong.

ABC News reported the same day that the standards are due by the end of 2026 and that Labor plans to legislate them in 2027. The speech itself does not give those dates, so treat them as the reported timetable until a bill is tabled.

1. 21 October 2025National AI Centre publishes the Guidance for AI Adoption and its six essential practices.
2. December 2025The National AI Plan relies on existing law and sector regulators. No AI Act.
3. 15 July 2026The Prime Minister announces the government will introduce AI laws.
4. September 2026PM&C consultation paper sets out the two groups: large data centres and frontier AI training in Australia.
5. 8 October 2026Andrew Charlton commits to legislating National AI Standards for frontier AI.
6. 9 October 2026Submissions on the consultation paper close.
7. 10 December 2026Privacy Act transparency on automated decisions commences. This one is already law and does apply to you.
8. End of 2026, as ABC reportedNational AI Standards due.
9. 2027, as ABC reportedLegislation planned. No bill had been introduced as at 12 October 2026.

Dashed dates come from ABC News reporting. The government has not announced them. Sources: National AI Centre, PM&C consultation paper, the minister's 8 October speech, ABC News 8 October 2026, OAIC.

Why now?

The speech opens with the reason. Last month the Prime Minister revealed that an AI agent built by a frontier lab had broken into an Australian Government system while it was being tested. It had been asked to look up public information about health spending. When it hit a barrier, it found a way around it, then did the same again. ABC reported that the system was a Medicare statistics portal. The minister's point: the lab did not catch it.

Who do the standards cover?

Two groups, and neither of them is a business that buys AI.

The detail is in the Department of the Prime Minister and Cabinet's consultation paper, Getting it right: Building AI infrastructure that works for Australia (September 2026, submissions closed 9 October 2026, PDF). It follows the Prime Minister's announcement on 15 July 2026 that the government will introduce AI laws, and it says the standards are "proposed to cover two related but distinct issues":

1. Large data centres. Mandatory minimum requirements on energy, water, costs, community engagement, location and workforce.
2. Frontier AI training in Australia. Conditions on the labs doing it.

For the second group, the paper sets out four areas:

• Security and safety. "Frontier labs who are granted authorisation to undertake large scale AI training in Australia will be required to adhere to specific, minimum-security and safety expectations, such as by disclosing defined reportable AI incidents to relevant Australian authorities."
• Information sharing. It asks whether developers should have to give information to the government up front, on request, or by publishing it.
• Local capability. Contributions to Australian research, skills and compute access.
• Creators and copyright. How Australian writers, musicians, artists and journalists get control and payment when their work is used in training.

The speech points the same way, at "developers of models with the sharpest, most acute frontier AI risks". Nothing in either document reaches a business that subscribes to Claude, ChatGPT or Copilot, or one that has an AI agent built on top of HubSpot.

Applies, or proposed to applyNot settled yet

Our reading of the September 2026 PM&C consultation paper and the 8 October speech. Neither says how obligations reach overseas models or the businesses that buy AI, so those links are dashed.

Here is how it splits:

What could change for buyers later?

This part is our reading, so it is labelled as such.

Your vendor's vendor may carry more paperwork. The paper's conditions attach to labs training models in Australia, so whether they reach the models you use depends on where those models are trained. Either way, once a lab has to show a working safety process and report incidents somewhere, you should be able to ask the business reselling or building on its models what that process says, and get an answer.

Other ministers are moving on the harms you can cause. The speech lists them: the Attorney-General on privacy reform and automated decision making, the Minister for Communications on a digital duty of care, the workplace minister on AI safety at work, and the Assistant Minister for Competition and Consumers on agentic commerce and surveillance pricing. Those are the areas most likely to reach a business using AI, and none of them waits for the frontier standards.

The systems model tends to travel. Workplace safety, CPS 230 and critical infrastructure rules all started with a defined group and became the expected standard for everyone they do business with. If you sell to banks, insurers or government, expect their procurement teams to ask you the same questions the regulator asks the labs. Our post on whether your AI vendor is a material service provider under CPS 230 covers the bank side.

What we do not know yet: whether any obligation flows down to businesses that deploy frontier models, how models trained overseas will be treated, and which agency will administer the frontier conditions and with what penalties. Neither the consultation paper nor the speech answers those.

Which framework applies to you today?

The National AI Centre's Guidance for AI Adoption, published 21 October 2025, replaced the 2024 Voluntary AI Safety Standard. It sets out six essential practices:

1. Decide who is accountable
2. Understand impacts and plan accordingly
3. Measure and manage risks
4. Share essential information
5. Test and monitor
6. Maintain human control

It comes in two versions. Foundations is for businesses earlier in their AI use. Implementation practices is for teams that build or customise AI, or use it in higher-risk ways such as hiring or customer decisions. The implementation version is explicit that accountability includes "contractors and third‑party providers/systems", and it lists supply chain accountability as an organisation-wide practice.

That is the closest thing a buyer has to the frontier standard today. It is voluntary, but it is what a regulator, a court or a big customer will measure you against if something goes wrong. If you are not sure which version you sit under, start with which Australian AI framework applies to you.

Eight questions to ask your AI vendor now

The minister's test for the labs is whether there is "a serious system in place to detect risks and prevent incidents occurring". Put the same test to whoever sells or builds your AI:

1. Which model, which version, and where does it run? Name the provider and the region. For Australian data, ask whether it stays in AWS Sydney (ap-southeast-2) or the equivalent.
2. What happens when the model provider changes the model? Who tests the change before your customers see it, and how long do you get?
3. What does your incident process look like, and when would we hear from you? Get a time in hours. If personal information is involved, the Notifiable Data Breaches clock is yours, so their notice to you has to come quickly.
4. Can the AI act, or only suggest? List every system it can write to. An agent that can change records needs a different conversation from a chatbot that answers questions.
5. Where is the human, and what can they stop? Ask to see the approval step, not hear about it.
6. What is logged, for how long, and can we get it? If you ever have to explain an automated decision to a customer under the Privacy Act changes, you need the record.
7. Is our data used to train anything? Get the answer in the contract, with the clause number.
8. Who owns what you built for us, and what happens if we leave? Under Australian law a contractor keeps copyright in commissioned code unless the contract assigns it. Our post on who owns the AI system after we leave goes through the clauses.

If those questions look familiar, it is because your largest customers are about to send them to you. The AI security questionnaire your buyer's CTO is about to send has the full list, and does the Essential Eight cover AI covers the security controls underneath.

What should a 100-person business do this quarter?

Four things, none of which waits for a bill:

• Write down every AI tool in use, including the ones staff signed up to themselves. That is your AI register, and the National AI Centre guidance expects one.
• Name one accountable person for each tool that touches customers or staff.
• Check your automated decisions against the Privacy Act change commencing 10 December 2026. Our post on the 10 December Privacy Act change explains what counts.
• Send the eight questions to your vendors and file the answers. If you are choosing a new one, use them alongside how to choose an AI implementation partner in Australia.

Frequently asked questions

Do Australia's new AI safety standards apply to my business? On what has been published, no, unless you run a large data centre or train frontier AI models in Australia. Those are the two groups the September consultation paper covers. Your business is still covered by the Privacy Act, Australian Consumer Law, work health and safety law and anti-discrimination law whenever it uses AI.

When do the national AI standards start? ABC News reported on 8 October 2026 that the standards are due by the end of 2026 and that legislation is planned for 2027. No bill had been introduced as at 12 October 2026.

What is systems regulation? It is a model where the government sets the standard a company's risk process has to meet, then holds the company accountable for whether that process works. Australia already uses it for workplace safety, bank supervision, critical infrastructure and parts of aviation.

Is the Guidance for AI Adoption mandatory? No. The National AI Centre's six essential practices are voluntary. They are the current government benchmark for responsible AI use and replaced the 2024 Voluntary AI Safety Standard on 21 October 2025.

Does the AI Safety Institute regulate businesses? No. Based on the government's description, it analyses how frontier AI is developing, advises on methods to keep it safe and supports agencies and regulators. It does not enforce rules on businesses that use AI.

Sources

• Speech: The Sydney Trust and Safety Festival, Assistant Minister Andrew Charlton, 8 October 2026, for every quote from the speech.
• Australia outlines plans to make AI companies prove their safety systems work, ABC News, 8 October 2026, for the reported 2026 and 2027 timetable.
• Guidance for AI Adoption, National AI Centre, for the six essential practices and supply chain expectations.
• PM&C AI infrastructure consultation summary, Licentium, for the September 2026 consultation paper proposals.
• Getting it right: Building AI infrastructure that works for Australia, Department of the Prime Minister and Cabinet, September 2026, for the two groups the standards cover and the frontier training conditions.

---

The government's question for the labs is whether there is a serious system behind the product. If your biggest customer asked you the same thing about the AI in your business tomorrow, what would you show them?

Want a second opinion on the AI tools your business already runs? See our AI consulting work, or talk to us.