Many Australian organisations built AI governance around the ten guardrails of the 2024 Voluntary AI Safety Standard, but that standard has been replaced and the planned mandatory guardrails for high-risk AI are not going ahead for now. The Department of Industry, Science and Resources (DISR) published the Voluntary AI Safety Standard on 5 September 2024 and, on 21 October 2025, replaced it with the Guidance for AI Adoption, which sets out six essential practices. Its consultation page on mandatory guardrails now says the government will not proceed at this time with those proposals. This guide explains what the current guidance asks for, how the ten guardrails map to the six practices, what is and is not binding, and what it means for AI agents and the machine identities they use. It is written for governance, risk, security and AI platform teams operating in Australia. For other jurisdictions see the Global AI Regulation Map.
Status line (checked 27 October 2026 on industry.gov.au and consult.industry.gov.au): the Guidance for AI Adoption is voluntary. There is no AI-specific statute in force. Existing laws continue to apply to AI use. Government policy may change, and the consultation wording "at this time" signals that mandatory rules could return; check DISR before relying on this summary.
Key takeaways
- The ten guardrails are superseded. DISR states that the Guidance for AI Adoption replaces the Voluntary AI Safety Standard and streamlines its ten guardrails into six practices.
- Mandatory guardrails are on hold. The 2024 proposals paper on mandatory guardrails for AI in high-risk settings will not proceed at this time, and DISR says the feedback informed the National AI Plan.
- Two versions for two audiences. Foundations is for organisations starting with AI; Extended practices give technical detail for organisations scaling AI.
- Accountability is personal. The first practice asks for a senior AI governance owner and a named accountable person for every AI system, including procured and embedded ones.
- Agentic AI is in scope. The risk practice says risk management should account for how agentic AI systems differ from other AI, and the human control practice asks for override points to pause, roll back or shut down systems.
Scope: who and what is covered
The Guidance for AI Adoption is aimed at businesses and organisations adopting AI, both technical teams and non-technical decision makers. It is not limited to high-risk uses or to developers: deployers that buy AI tools, or use AI embedded in other software, are expected to apply it too. DISR says it stays aligned with Australia's AI Ethics Principles and international standards.
| Document | Date | Status at 27 October 2026 |
|---|---|---|
| Voluntary AI Safety Standard (10 guardrails) | Published 5 September 2024 | Superseded by the Guidance for AI Adoption |
| Proposals paper: Introducing mandatory guardrails for AI in high-risk settings | Consultation 5 September to 4 October 2024 | Government will not proceed at this time; feedback informed the National AI Plan |
| Guidance for AI Adoption (6 essential practices), Foundations and Extended practices | 21 October 2025 | Current voluntary guidance |
The supporting tools published with the guidance are an AI screening tool (to decide how much governance oversight each AI system needs), an AI policy guide and template, an AI register template, and a terms and definitions list.
Key obligations
None of these are legal obligations. They are the practices DISR expects responsible organisations to adopt. The actions below are from the Foundations version.
| Essential practice | Key actions | Identity and security angle |
|---|---|---|
| 1. Decide who is accountable | Senior AI governance owner; AI policy; accountable person per AI system; training; supply chain responsibilities; governance framework | Every AI system and agent has a named human owner in the register |
| 2. Understand impacts and plan accordingly | Stakeholder impact assessment with attention to vulnerable groups; contestability channels; stakeholder engagement; monitor feedback patterns | Contest routes for AI decisions about customers or staff |
| 3. Measure and manage risks | Risk screening to flag unacceptable or high-attention uses; risk management that accounts for different AI types, including agentic AI; use-case risk assessments; proportionate controls; investigate and learn from AI incidents | Agent permissions and tool access treated as risk factors |
| 4. Share essential information | AI register including procured and embedded AI; disclose AI use, especially where AI affects decisions or could be mistaken for a human; document capabilities and limitations; supply chain transparency; explain outcomes in proportion to their seriousness | Register entries should record the system's credentials and integrations |
| 5. Test and monitor | Ask suppliers for proof of testing; test before deployment; monitor after deployment; extend data governance and cybersecurity practices to AI; stress test and consider independent testing for higher-risk uses | Security testing of AI includes prompt injection and credential exposure |
| 6. Maintain human control | Oversight matched to autonomy and stakes; override points to pause, override, roll back or shut down; train overseers; alternative pathways if AI fails or is retired | Credential revocation is the practical shutdown for an agent |
How the ten guardrails map to the six practices
DISR says the ten guardrails were streamlined into six practices but does not publish a guardrail-by-guardrail crosswalk on the pages checked. The mapping below is NHI Mgmt Group's own reading, to help organisations that already have guardrail-based controls.
| 2024 guardrail (condensed label) | Closest 2025 practice |
|---|---|
| 1. Accountability and governance | 1. Decide who is accountable |
| 2. Risk management | 3. Measure and manage risks |
| 3. Data governance and protection | 5. Test and monitor (extend data governance and cybersecurity to AI) |
| 4. Testing and monitoring | 5. Test and monitor |
| 5. Human control and oversight | 6. Maintain human control |
| 6. Transparency to end users | 4. Share essential information |
| 7. Contestability | 2. Understand impacts and plan accordingly |
| 8. Supply chain transparency | 1 and 4 (supply chain accountability and transparency) |
| 9. Records and documentation | 4. Share essential information (AI register, documentation) |
| 10. Stakeholder engagement | 2. Understand impacts and plan accordingly |
Timeline and current status
- 5 September 2024: Voluntary AI Safety Standard published, with ten voluntary guardrails. The proposals paper on mandatory guardrails for high-risk AI opened for consultation the same day and closed on 4 October 2024.
- 21 October 2025: Guidance for AI Adoption published, replacing the Voluntary AI Safety Standard.
- Current: the consultation page states that the government will not proceed at this time with previous proposals to introduce mandatory guardrails, and that feedback informed the National AI Plan. The National AI Plan page itself could not be retrieved for this guide, so its contents are not summarised here.
"At this time" leaves the door open. Organisations that align with the six practices now will be better placed if binding rules return, and many of the same controls are required by other regimes they may already face, such as the EU AI Act. See the EU AI Act Compliance Guide.
Enforcement and penalties
The Guidance for AI Adoption has no enforcement mechanism or penalties. Legal risk from AI use in Australia comes from laws that already apply regardless of technology, including privacy, consumer protection, anti-discrimination, sector regulation and cyber security law. Two NHI Mgmt Group guides cover the most relevant ones:
- Privacy: the Privacy Act 1988 and the Privacy and Other Legislation Amendment Act 2024, which added a transparency duty for automated decisions. See the Australia Privacy Act Reforms Guide and the Australian Privacy and Cyber Rules Guide.
- Cyber security and critical infrastructure: the Australia Cyber Security Act 2024 Guide and the Australia SOCI Act Guide cover duties that apply when AI systems are part of critical or connected infrastructure.
Government agencies may have separate mandatory AI policies set by central agencies; those are outside this guide.
What it means for AI agents and non-human identities
The guidance names agentic AI in its risk practice and its human control practice is written for systems that act. Put into identity terms:
- Register agents as AI systems: practice 4's AI register should list every agent, including those embedded in SaaS products, with its accountable person (practice 1), purpose and risk rating.
- Give each agent its own identity: distinct credentials per agent make the accountable person meaningful and allow scoped permissions and audit. See the Agent Identity and Access Guide.
- Scope tool access by risk: practice 3 asks for proportionate controls. Read-only agents and agents that can send payments or change records are different risk classes.
- Build the off switch into the credential: practice 6 asks for points to pause, override, roll back or shut down. For an agent, that means short-lived tokens, a tested revocation path and approval gates for high-impact actions.
- Disclose agents to people: practice 4 asks organisations to disclose AI use where it could be mistaken for a human, which covers chat, voice and email agents.
- Test for agent-specific failure: practice 5 testing should include prompt injection, excessive permissions and credential leakage. See the Red Teaming AI Agents for Identity Abuse guide.
Compliance steps
- Re-baseline: if your AI governance references the ten guardrails, map it to the six practices using the table above and update policy references.
- Appoint the owner: name a senior AI governance owner and an accountable person per AI system.
- Build the register: start from the DISR AI register template; include procured, embedded and agentic AI, and record credentials and integrations.
- Screen and assess: use the AI screening tool to decide oversight levels, then run use-case risk and impact assessments for higher-attention systems.
- Set up contestability and disclosure: publish how people can challenge AI decisions and where AI use is disclosed.
- Test and monitor: collect supplier testing evidence, test before deployment and monitor in production; extend existing security controls to AI.
- Prove human control: document and rehearse override and shutdown for each system, including agent credential revocation.
- Choose the version: small organisations can work from Foundations; organisations scaling AI should adopt Extended practices for the technical detail.
Foundations or Extended practices?
| Situation | Start with | Why |
|---|---|---|
| Small business using off-the-shelf AI tools (chat assistants, AI features in SaaS) | Foundations, AI policy template, AI register template | Most duties are about ownership, disclosure and knowing what AI you use |
| Organisation building AI into customer products or decisions | Extended practices for the systems concerned | Needs testing, monitoring, contestability and documentation at technical depth |
| Organisation deploying AI agents with access to business systems | Extended practices, with the screening tool used per agent | Autonomy and tool access raise the oversight level the screening tool should assign |
| Multinational already aligned to EU AI Act, NIST AI RMF or ISO/IEC 42001 | Map existing controls to the six practices | Avoids a parallel programme; gaps are usually in disclosure and contestability |
Common pitfalls
- Policies still citing the ten guardrails: auditors and customers will ask why governance references a superseded standard. Update references and keep the mapping as evidence of continuity.
- Registers that miss embedded AI: the guidance explicitly includes procured and embedded AI. AI features switched on inside SaaS products, and agents connected through OAuth grants, are the usual gaps. See the Shadow AI Discovery Guide.
- Accountability without authority: a named accountable person who cannot stop or change the system does not meet the intent of practices 1 and 6.
- Shutdown that has never been tested: an override point that exists only on paper fails when an agent misbehaves at speed. Rehearse revoking agent credentials and disabling integrations.
- Treating "voluntary" as optional for regulated activity: privacy, consumer and sector rules still apply to what the AI does, whatever the status of the guidance.
Mapping to other frameworks
| Practice | NIST CSF 2.0 | NIST SP 800-53 Rev. 5 |
|---|---|---|
| Decide who is accountable | GV.RR-02 roles and responsibilities | PM-2 |
| Measure and manage risks | ID.RA-01, GV.RM | RA-3 |
| Share essential information (register) | ID.AM-02 software and services inventoried | CM-8 |
| Test and monitor | DE.CM-09 | SA-11, SI-4 |
| Maintain human control | PR.AA-05 access permissions | AC-6, IR-4 |
Metrics: share of AI systems in the register with an accountable person (target 100%); share of agents with distinct credentials and a tested revocation path; number of AI systems screened and assessed; AI incidents logged and reviewed per quarter; contest requests received and median resolution time.
Practitioner checklist
- Replace references to the Voluntary AI Safety Standard with the six essential practices.
- Name a senior AI governance owner and an accountable person per AI system.
- Build an AI register covering procured, embedded and agentic AI, with credentials and integrations.
- Screen every AI use case and assess higher-attention ones before deployment.
- Publish contest routes and disclose AI where it could be mistaken for a human.
- Collect supplier testing evidence and monitor systems after deployment.
- Give each agent its own identity, scoped permissions and a tested shutdown path.
- Track Privacy Act automated decision duties and watch DISR for policy changes.
- Rehearse shutdown of at least one agent per quarter and record the time taken to revoke its access.
Standards and references
- Department of Industry, Science and Resources: Guidance for AI Adoption
- Department of Industry, Science and Resources: Guidance for AI Adoption: Foundations
- Department of Industry, Science and Resources: Voluntary AI Safety Standard
- Department of Industry, Science and Resources: The 10 guardrails
- Department of Industry, Science and Resources: Introducing mandatory guardrails for AI in high-risk settings
- NIST: Cybersecurity Framework 2.0
- NIST: SP 800-53 Rev. 5
This guide summarises Australian Government AI guidance as at 27 October 2026 and is not legal advice. Related NHI Mgmt Group resources: Global AI Regulation Map · Australia Privacy Act Reforms Guide · Agent Identity and Access Guide · NIST AI RMF Guide