Independent source-linked navigator Β· Includes 1 May 2026 Agentic AI update

Australia's AI Governance
8 Principles. 6 Practices.
One Navigator.

The most common confusion in Australian AI governance β€” decoded. Understand what each layer means, how they fit together, and how the 1 May 2026 agentic AI update strengthens safe adoption.

Why & How.
Two different questions.

Australian AI guidance commonly answers two different questions. The 8 Ethics Principles answer why β€” they describe the responsible AI values published in 2019.

The 6 AI Essential Practices (AI6) answer how β€” they provide practical steps for responsible AI governance and adoption, published in October 2025 as an update to VAISS.

They are complementary, not competing. AI6 maintains alignment with the 8 principles, so the page shows them together instead of treating them as separate documents.

8
AI Ethics Principles
The why β€” responsible AI values (2019) Β· Source: DISR
↓ built upon ↓
6
AI Essential Practices (AI6)
The how β€” practical adoption guidance (October 2025) Β· Source: NAIC
What happened to the 10 guardrails?
The Voluntary AI Safety Standard (VAISS) had 10 guardrails (2024). NAIC condensed 10 guardrails into 6 essential practices in October 2025. Use the 6-practice structure as the simpler current view, while checking the source documents for exact wording.

How the numbers evolved

2019
Australia's 8 AI Ethics Principles published
What changed for practitioners: responsible AI moved from β€œnice idea” to a common values baseline β€” fairness, privacy, transparency, safety and accountability became the language for reviews.
Still current Β· Foundational
2024
Voluntary AI Safety Standard β€” 10 Guardrails
What changed for practitioners: organisations gained a more concrete checklist for AI governance, risk, transparency, testing and human oversight.
Superseded Oct 2025
2025
AI6 β€” 6 Essential Practices released
What changed for practitioners: the old 10-guardrail view became a simpler 6-practice operating model for owners, risk teams, delivery teams and assurance reviewers.
Current source Β· AI6
2026
Agentic AI security guidance published
What changed for practitioners: autonomous agents need tighter identity, access, tool control, monitoring, incident response and human approval than simple chat assistants.
Current Β· Agentic AI

The 8 AI Ethics Principles

Click a principle. The detail opens beside the cards, so the full context stays visible on one screen.

1
Human, Societal & Environmental Wellbeing
Benefit people, society and environment.
2
Human-Centred Values
Respect rights, diversity and autonomy.
3
Fairness
Avoid unfair discrimination and exclusion.
4
Privacy Protection & Security
Protect data and secure the system.
5
Reliability & Safety
Test, monitor and manage safety risks.
6
Transparency & Explainability
Explain when and how AI affects people.
7
Contestability
Provide a way to challenge outcomes.
8
Accountability
Make accountable owners identifiable.

The 6 Essential Practices (AI6)

Click a practice. The practical steps open beside the cards, so users can see the control and the explanation together.

1
Decide Who Is Accountable
Name owners and decision responsibilities.
2
Understand Impacts and Plan Accordingly
Assess affected people, harms and context.
3
Measure and Manage Risks
Manage AI risk through enterprise controls.
4
Share Essential Information
Explain use, purpose, data and impacts.
5
Test and Monitor
Test before release and monitor production.
6
Maintain Human Control
Keep review, override and challenge paths.

Agentic AI is the new risk layer

This is not presented here as a separate governance framework. It is a security and control lens for AI systems that can reason, plan, use tools, access data, call APIs and act with some autonomy. Use it to strengthen the 8 Principles and AI6 practices already shown on this page.

πŸ›‘

Source update Β· 1 May 2026: ASD's ACSC, CISA, NSA, the Canadian Centre for Cyber Security, NCSC-NZ and NCSC-UK published a joint source document called Careful adoption of agentic AI services.

The practical message for this navigator: start with low-risk and non-sensitive tasks, avoid broad access, and design for identity, least privilege, monitoring, auditability, human oversight and rollback.

Normal GenAI

Creates content for humans

Standard generative AI mainly produces text, images, summaries, drafts or analysis. A human usually decides what to do next.

  • Lower autonomy by default
  • Usually limited to advisory output
  • Main risks: privacy, hallucination, bias, data leakage and weak review
Agentic AI

Can take action through tools

Agentic AI builds on GenAI by integrating with software systems so it can reason, plan and take actions without continuous human intervention.

  • Higher autonomy and operational impact
  • Can call APIs, use tools, access memory and trigger workflows
  • Main risks: excessive privilege, prompt injection, tool misuse, cascading failure, weak accountability
πŸ”
Before build: access
Limit identity, data and tool permissions before the agent is connected to real systems.
🧩
During build: tool chain
Check prompts, memory, APIs, plugins and handoffs as one combined attack surface.
πŸ§ͺ
Before go-live: failure tests
Test prompt injection, unsafe actions, bad instructions, overreach and human override paths.
🧾
In operation: accountability
Log what the agent saw, decided and did so incidents can be investigated and explained.

Agent-safe adoption checklist

Use this as the simple enterprise pattern: do not block innovation, but do not allow agents to become invisible privileged users. Every agent should have a named owner, bounded purpose, limited permissions, logging, monitoring, human approval points and rollback.

How this maps back to the existing framework

This keeps your site message clean: the 8 Principles are still the values, AI6 is still the operating model, and the 1 May 2026 agentic AI guidance is the extra security pattern for autonomous tool-using systems.

Agentic AI control AI6 alignment Ethics principle alignment Plain-English governance test
Named agent ownerPractice 1 β€” Decide who is accountableP8 AccountabilityCan we name the person accountable for this agent's actions?
Agent risk assessmentPractice 2 β€” Understand impactsP1 Wellbeing, P3 Fairness, P4 Privacy/SecurityHave we assessed who can be harmed and how?
Least privilege and JIT accessPractice 3 β€” Manage risksP4 Privacy/Security, P8 AccountabilityCan the agent only access what it needs, when it needs it?
Tool, memory and action logsPractice 4 β€” Share essential informationP6 Transparency, P8 AccountabilityCan we explain what the agent saw, used, decided and did?
Testing, red teaming and monitoringPractice 5 β€” Test and monitorP5 Reliability/SafetyCan we detect drift, misuse, unsafe actions or degraded performance?
Human approval for high-impact actionsPractice 6 β€” Maintain human controlP2 Human-centred values, P7 ContestabilityCan a human stop, challenge, reverse or approve high-impact actions?
Source anchor: ASD ACSC joint guidance, Careful adoption of agentic AI services, first published 1 May 2026. This section is a practical interpretation for enterprise governance and is not legal advice.


Move from idea to production without skipping guardrails

A practical train-style route for AI use cases and agentic AI. Each station is a gate before the use case moves forward.

AI adoption route

Start small. Control access. Test hard. Monitor continuously.

Use this route for normal AI use cases and add stronger controls where the AI agent can use tools, call systems, access memory or trigger workflows.

1DiscoverPurpose and value
2GovernOwner and risk
3ControlData, tools, access
4AssureTest and approve
5OperateMonitor and retire
01

Choose the use case

Define the outcome, users, value and boundaries. Confirm AI is needed.

AI6 P2Discover
02

Name the owner

Assign who approves, operates, monitors and responds to issues.

AI6 P1Accountability
03

Assess impact

Check privacy, fairness, safety, security, legal and people impact.

AI6 P2/P3Risk
04

Control access

Limit data, APIs, memory, actions and tool permissions.

Agent gateLeast privilege
05

Test hard

Test edge cases, unsafe outputs, prompt attacks and tool misuse.

AI6 P5Red team
06

Approve release

Use register entry, rollback, user guidance and human approval points.

AI6 P4/P6Human control
07

Monitor live use

Track logs, sources, drift, incidents, complaints and business value.

OperateMonitor
08

Retire or redesign

Remove access, archive records and redesign if value or risk changes.

RetireRecords
Simple rule: agents need extra gates at access, testing and release because they can act through tools. Do not give them broad access and hope governance will catch up later.

A few practical test cases to make the guidance real

Use these examples as patterns. They are not legal or compliance decisions β€” they show how to turn the 8 Principles, AI6 and agent controls into action.

Agentic AI example

Employee onboarding agent

An internal agent answers onboarding questions, searches HR/IT policy, and raises low-risk service requests for new starters.

Primary risk: the agent may access more employee information or systems than needed, or trigger actions without enough human control.

1. Classify the use case

Agentic AI, internal-facing, medium risk if it can create tickets or call workflows.

2. Apply AI6

  • Name HR and IT owners before build.
  • Record purpose, data sources and tool access.
  • Test answers against approved onboarding policies.

3. Apply ethics principles

  • Privacy and security: limit personal data access.
  • Transparency: tell users they are using an AI agent.
  • Accountability: define who handles errors and complaints.

4. Agent gate

  • Start read-only where possible.
  • Use least privilege for tools and workflows.
  • Require human approval for high-impact account or HR actions.
Decision outcome Pilot with limited users, read-only knowledge access first, and manual approval for any workflow action.
Check the lifecycle route β†’
Low-risk pattern Internal summarisation of public or approved internal content. Main action: owner + source control + review.
Medium-risk pattern Assistant that uses personal information or creates tickets. Main action: privacy check + testing + monitoring.
High-risk pattern AI influencing rights, eligibility, employment, finance, health or access to services. Main action: formal review + human control + contestability.
Agentic pattern AI that can use tools, call APIs or trigger workflows. Main action: least privilege + tool logging + rollback.

Frequently asked

Are the 8 Ethics Principles still useful if AI6 updated VAISS?
Yes. The 8 Ethics Principles are still a useful values layer. The October 2025 Guidance for AI Adoption says the 6 practices maintain alignment with Australia’s AI Ethics Principles and international standards. This page shows both layers together so readers can move from values to practical controls without jumping between documents.
How should I treat AI6?
For general business use, the Guidance for AI Adoption is presented as practical guidance for responsible AI adoption. Government agencies and regulated sectors may also have separate policy, privacy, cyber, procurement or sector-specific obligations. Use AI6 as a clear operating model, but check the source pages and your organisation’s legal/compliance requirements before treating it as a binding requirement.
If my organisation used VAISS, do I need to start over?
No, not from zero. The October 2025 update condensed the 10 VAISS guardrails into 6 essential practices. Use the 6-practice view to simplify your current register, controls and review process. Then check the source documents for any detailed control gaps, especially around risk, monitoring, transparency and human control.
Which principle covers the Privacy Act APP 1.7 automated decision-making obligation?
Principle 6 β€” Transparency & Explainability is the primary ethics anchor. People must understand when AI significantly impacts them. The incoming APP 1.7–1.9 privacy policy obligations (from 10 December 2026) operationalise this principle by requiring all APP entities to disclose in their privacy policy: the types of personal information used in substantially automated decisions, and the nature of decisions that could significantly affect individual rights. This is also covered by AI6 Practice 4 β€” Share Essential Information. Begin privacy policy updates now β€” the obligation applies to all decisions made from that date, regardless of when the AI system was deployed.
Does Principle 7 (Contestability) require a formal appeal process?
In practical terms, yes for high-risk use cases. Principle 7 states that a "timely process" must allow people to challenge AI use or outcomes when they are significantly affected. AI6 Practice 6 (Maintain Human Control) operationalises this by requiring challenge and appeal pathways to be designed before deployment. Privacy Act Tranche 2 reforms (expected post-2027) are expected to formalise a statutory right to request meaningful information about automated decisions. Don't wait β€” build contestability into your AI systems now as leading practice. Document the appeal pathway in your AI register.
Should older AI systems be reviewed?
Yes. Even if a system was built earlier, active AI use should be reviewed against current governance, privacy, cyber and business risk expectations. For privacy, OAIC guidance says APP entities will have extra privacy policy obligations from 10 December 2026 where personal information is used by a computer program for decisions that could significantly affect an individual’s rights or interests.
Where does the 1 May 2026 agentic AI guidance fit?
It sits as a security and operational risk layer on top of the existing framework. The 8 Principles still explain the values, AI6 still explains the responsible AI operating model, and the ACSC agentic AI guidance explains what changes when an AI system can use tools, access data, call APIs, create sub-tasks or act without continuous human intervention. In practice, agentic AI needs stronger identity, least privilege, monitoring, logging, human approval points, safe defaults, isolation and rollback.
How does this relate to the DTA AI Technical Standard?
The DTA AI Technical Standard is a separate source for government AI delivery and technical governance. This page does not replace that standard. It simply helps readers understand the relationship between values, adoption practices and the newer agentic AI control lens. Source: digital.gov.au/policy/ai/AI-technical-standard.

Go to the source

Short links to the public source pages used to build this independent navigator.