Australia's new AI safety rules: what they mean for your app

On 10 November 2026 the Government announced new AI safety rules: mandatory safety and security protections, enforced by a new regulator. If your app calls an AI model, the question is whether any of that is now your problem. Mostly it is not, at least not directly. The duties fall on the companies that build the most advanced models. What reaches your app comes through rules that already apply, and several of them changed this year.

What the AI safety rules require

The protections apply to frontier AI companies that want to operate in Australia. The Government describes frontier AI as the most advanced models at the current cutting edge. Under the proposal, those companies must:

• keep security controls across development and deployment
• identify and manage serious risks from their models
• allow independent evaluation of their models
• notify the Government of incidents promptly
• maintain and publish safety and security frameworks
• adapt as standards and rules are updated.

A new regulator would enforce them alongside the Australian AI Safety Institute, with what the release calls significant penalties. The Government is also considering changes to criminal law for serious misconduct. It will consult further before it publishes the legislation, and it has not yet set the threshold that makes a company "frontier".

This sits inside the Australian Standards for AI, the mandatory framework the Prime Minister announced in July 2026, with legislation expected in 2027. Until now the published detail of that framework covered large AI data centres and the use of creative work to train models.

Who it covers, and who it doesn't

If your app sends text to OpenAI, Anthropic or Google through their API, you are their customer. You are not building a frontier model, and the new duties are not yours.

The announcement still touches you in two ways.

Your supplier will have to publish more. A frontier provider that must publish a safety and security framework gives you something to cite when a buyer asks how your AI supplier manages risk. Today most founders answer that question with a link to a marketing page.

Your model will keep changing. Providers already update models without asking. Under a duty to adapt to new standards, expect more of it. If your app calls a "latest" model alias, its behaviour can change between one day and the next with no release on your side. Pinning the model version, and having a test that would notice a change, is how you decide when that happens.

The rules that reach your app today

The Privacy Act

The Act has no rule that data must stay in Australia. Under APP 8, it makes you answerable for personal information you send overseas, and user text sent to a hosted model usually counts. Our post on data residency for AI-built apps covers that in detail.

The newer change starts on 10 December 2026. If a computer program uses personal information to make, or substantially and directly help make, a decision that could significantly affect someone's rights or interests, your privacy policy must say so. The OAIC has published guidance on the test. An AI feature that screens applicants, sets prices or decides eligibility should be checked against it.

Apple's guideline 5.1.2(i)

Since November 2025, Apple's App Review Guidelines require apps to "clearly disclose where personal data will be shared with third parties, including with third-party AI, and obtain explicit permission before doing so." A line in the privacy policy is not explicit permission. The app has to ask, before the first request goes to the model.

The eSafety industry codes

The Phase 2 industry codes under the Online Safety Act have been in force since 9 March 2026, and they reach generative AI services as well as social media. Services with an AI companion chatbot feature carry their own risk assessment for harmful content, and protections for children. Whether your assistant counts as a companion is eSafety's call, and it turns on what the assistant is designed to do. Read the registered codes if your app talks to the general public, teenagers included.

The TGA, if your app touches health

An AI coach that suggests a treatment change or a dose is far more likely to be regulated as a medical device than one that records how someone feels. Classification turns on intended use, and the wording in your store listing is part of how intended use is read.

If you sell to banks, insurers or super funds

APRA's letter to industry on AI of 30 April 2026 went to the entities it regulates, not to you, and it sets no new rules. Even so, APRA asks them to "maintain visibility over the full AI supply chain", to get "transparency, auditability and assurance" from AI providers through their contracts, and to security-test AI-generated code. The letter also prefers "enforceable technical restrictions or robust preventative controls" to policy.

Your product is part of that supply chain, so those expectations arrive in the security questionnaire. A claim that your assistant "never sends data off the device" or "can't take actions" will be read as a technical restriction, and the reviewer will want to know how it is enforced. A rule written into a system prompt is a policy.

What to check in your own app

Most of this you can answer yourself:

• List every AI model the app calls, the provider, the region and what user data goes in each request.
• Check that the app asks permission before the first request that carries personal data.
• Pin the model version, and keep a test that would catch a change in behaviour.
• Decide what the app does when the provider is down, and try it.
• If the AI can call tools or act for the user, write down what each tool can reach and which actions need a person to approve them.
• Read your privacy policy against the 10 December automated-decision test.

None of this is legal advice. Where a classification or a disclosure matters, ask a lawyer who works in that area.

If you want to know which gates your app faces, the free Launch Gate Check names each one with its guideline number. The Envisage Apps Launch Gate Audit tests them on your build. For an app with AI in it, that includes which providers receive your users' data, Apple's consent rule, what an agent is allowed to do, and the technical questions a bank's reviewer will ask. A senior engineer validates every finding.