The Voluntary AI Safety Standard has been superseded: what Guidance for AI Adoption means for your organisation

Most Australian organisations that wrote an AI policy in the last two years wrote it against the Voluntary AI Safety Standard and its 10 guardrails. That standard has been superseded, and a policy that still cites it as its authority is now pointing at a document the Government has moved on from.

The replacement is not a reversal, and organisations that genuinely adopted the guardrails are not starting again. But the reference point has changed, the structure has been simplified, and the mapping in most AI policies is now stale. This article sets out what changed, what it means in practice, and what a board should ask.

Key takeaways

• The Voluntary AI Safety Standard, and its 10 guardrails, has been superseded. On 21 October 2025 the National AI Centre published Guidance for AI Adoption, which the Department of Industry, Science and Resources describes as updated and simplified guidance that evolves the standard.¹
• The 10 guardrails have become six essential practices: decide who is accountable; understand impacts and plan accordingly; measure and manage risks; share essential information; test and monitor; and maintain human control.¹
• The guidance now comes in two versions: a foundations version for organisations early in their AI use, and an implementation version for those building or customising AI, or managing higher-risk use cases.¹
• It remains voluntary and creates no new legal duties. It is still the benchmark an organisation is likely to be measured against if an AI system causes harm.
• It aligns with AS ISO/IEC 42001:2023, the AI management system standard Standards Australia adopted as an identical Australian Standard on 16 February 2024.² Designing to the standard, rather than to guidance that is still moving, is the more durable choice.

What was the Voluntary AI Safety Standard?

The Voluntary AI Safety Standard was the Department of Industry, Science and Resources’ practical guidance to Australian organisations on using AI safely and responsibly. It set out 10 guardrails covering accountability, risk management, data governance, testing, human oversight, transparency to users, contestability, supply chain information sharing, record-keeping and stakeholder engagement.³

It applied across the AI supply chain, to developers and deployers alike, and it was explicitly voluntary: it created no new legal obligations. Its purpose was to give organisations something concrete to adopt while the question of mandatory regulation for high-risk AI was still being worked through. The guardrails were updated in 2024 to align more closely with the Government’s proposed mandatory guardrails.³

For a lot of Australian organisations it did its job. It was the first document that translated “use AI responsibly” into activities someone could be made accountable for.

What replaced it?

On 21 October 2025 the National AI Centre published Guidance for AI Adoption. The Department describes it as updated and simplified guidance for industry that evolves the Voluntary AI Safety Standard.¹

Two things changed structurally.

Ten guardrails became six essential practices. The consolidation is mostly a tightening rather than a reduction in expectation. The activities that sat under the old guardrails have largely been folded into fewer, broader practices.

The guidance split into two versions. There is now a foundations version, for organisations earlier in their AI use, and an implementation version, aimed at teams that build or customise AI systems, use AI in more complex ways, manage higher-risk use cases, or need stronger controls and oversight.¹ This is a sensible acknowledgement that a council running a chatbot and a bank running credit models do not need the same document.

What are the six essential practices?

1. Decide who is accountable. Assign, document and communicate who is accountable for the AI management system, for each AI system’s development and deployment, for oversight of third-party AI, for testing, and for handling concerns and requests for redress. It also covers supply chain accountabilities and AI literacy.¹
2. Understand impacts and plan accordingly. AI operates at speed and scale, which magnifies its effects. A single system can produce widespread unfair decisions or inaccurate information before anyone notices.
3. Measure and manage risks. AI-specific risk management, applied to each use case and reviewed on an ongoing basis rather than once at go-live.
4. Share essential information. Transparency to users about AI-enabled decisions and AI-generated content, and to others in the supply chain about data, models and systems.
5. Test and monitor. Test before deployment against defined acceptance criteria, then monitor for behaviour change and unintended consequences.
6. Maintain human control. Meaningful human oversight, and the ability to intervene, across the AI lifecycle.

The guidance is direct about the unit of analysis: the same tool can create very different risks depending on how it is used. Its own example is that an internal chatbot carries different risks from AI used in hiring or customer decisions, and each use should be reviewed on its own.¹ This is the point most AI policies get wrong. They govern tools, when the risk lives in use cases.

Does voluntary guidance actually matter?

It is voluntary, it creates no new legal duties, and it is tempting to file it accordingly. That would be a mistake, for three reasons.

Existing law still applies. Nothing about AI suspends privacy obligations, work health and safety duties, anti-discrimination law, consumer law or directors’ duties. A decision made with a model is still the organisation’s decision. Voluntary guidance does not create the obligation; the obligation was already there.

It is the yardstick. When something goes wrong, the question asked afterwards is whether the organisation acted reasonably. Published government guidance on what reasonable practice looks like is the most likely answer to that question. Choosing not to follow it is a defensible position only if the organisation can explain what it did instead.

Sector obligations reach further. Australian Government entities carry Protective Security Policy Framework requirements that AI deployments engage directly, including policy advice on generative AI use with OFFICIAL information. Responsible entities under the Security of Critical Infrastructure Act must manage material risks to their assets, and AI embedded in an operational process is part of that hazard picture rather than separate from it.

What should an organisation do now?

If your AI policy cites the Voluntary AI Safety Standard, it needs a refresh rather than a rewrite. A short, practical sequence:

• Re-point the reference. Find every document citing the Voluntary AI Safety Standard or the 10 guardrails and map it to the six essential practices. In most cases the underlying controls survive intact.
• Check which version applies. Foundations or implementation. Getting this wrong in either direction is expensive: an organisation running higher-risk use cases against the foundations guidance is under-governed, and one running an internal drafting tool against the implementation guidance will exhaust its people on paperwork.
• Build the inventory, if you have not. The first practice is accountability, and accountability requires knowing what exists. Most organisations underestimate their AI footprint, because much of it arrived inside tools they already licensed rather than through a procurement decision.
• Move the unit of analysis to the use case. Assess per use case, not per tool.
• Anchor to the standard, not the guidance. Guidance has changed once already and will change again. AS ISO/IEC 42001:2023 is a published standard with a stable structure, and the guidance aligns to it. Designing the management system to the standard means the next guidance update is a mapping exercise rather than a rebuild.
• Put AI failure into the plans you already have. An AI system that fails in a process that matters is a resilience event. It belongs in business continuity and crisis arrangements, not in a separate AI document nobody opens.

Questions worth putting to your board

Four questions that tend to expose the gap quickly:

1. Where are we using AI, and who is named as accountable for each consequential use?
2. Which of our AI use cases could affect a person’s rights, safety, money or access to a service, and what human control exists over those?
3. If an AI system produced a harmful or wrong outcome tomorrow, what evidence could we produce about how it was tested, approved and monitored?
4. What have we told our suppliers we require of them, and what have they told us about their models?

If those cannot be answered without a round of emails, the gap is governance rather than technology.

How Agilient can assist

Agilient is an independent security, risk and resilience consultancy. We advise on the governance of AI, and we do not build, sell or resell it, so the advice is shaped by your obligations rather than by a product range.

For organisations working through the change from the Voluntary AI Safety Standard, the usual starting point is an AI governance maturity assessment: a scoped, desktop-first diagnostic against AS ISO/IEC 42001 and AS ISO 31000 that returns a maturity rating, a prioritised roadmap and a board-ready summary. From there, Agilient delivers AI risk assessments that write into your existing risk register, independent assurance reviews, AI procurement and vendor assurance, AI policy frameworks, board briefings, and scenario exercises for AI failure or misuse.

The method is the one Agilient already applies across security risk assessment, protective security and resilience work for government, critical infrastructure, health and regulated industry. AI risk extends that method rather than replacing it, which is also the cheapest way to govern it. Background on the standards and the current Australian guidance is maintained on the AI governance pillar.

Frequently asked questions