The Australian AI Data-Governance Index 2026

Methodology note

The analytical lens

We read the three datasets through a single question: is Australia's data-governance capability keeping pace with the speed of AI adoption in its most-regulated sectors? The OAIC data describes where Australian data is leaking and why. The ASD data describes the threat volume and the rising cost. The Responsible AI Index describes how mature the governance response is. Read together, they make the case for runtime AI data controls in regulated environments.

Caveats we hold honestly

• Correlation, not causation. Breach concentration in health, finance, and government reflects the volume and sensitivity of data those sectors hold, not a proven AI link. We use the overlap with fast AI adoption to frame governance urgency, not to attribute any breach to AI.
• Self-report bias. The ASD cost figures are self-reported by businesses² and the Responsible AI Index is a decision-maker survey³; both tend to skew optimistically on maturity and conservatively on cost. We surface the figures as the sources reported them.
• Half-year versus annual windows. The OAIC figure covers the first half of 2025¹; the ASD figure covers the full 2024-25 financial year². We do not annualise or combine the windows.
• Commercial incentive disclosure. Areebi sells a secure-AI-control-plane product whose data-loss-prevention layer is relevant to the human-error finding. We disclose this incentive up front; the underlying sources are linked so you can re-do the synthesis with different weights.

The Australian breach landscape

The OAIC's Notifiable Data Breaches Report for January to June 2025 recorded approximately 532 notifications, down around 10% on the previous half¹. The fall in volume is welcome, but the composition is the part that matters for AI governance: where the breaches land, and what causes them.

The three most-affected sectors - health at approximately 18%, finance at around 14%, and the Australian Government at around 13%¹ - are exactly the regulated sectors moving fastest into AI-assisted workflows: clinical documentation, claims and fraud triage, and citizen-service automation. The data those sectors hold is the most sensitive in the economy, and it is the same data that now flows toward AI systems. That overlap is the central governance concern of this Index.

The threat backdrop is intensifying. The ASD Annual Cyber Threat Report 2024-25 logged over 84,700 cybercrime reports - about one every six minutes - and responded to over 1,200 cyber security incidents, up approximately 11% on the prior year². A higher tempo of attack against the sectors holding the most sensitive data raises the stakes for every new AI data flow those sectors open.

For the regulatory treatment of breach obligations in Australia, see our Australian Privacy Act compliance brief and the companion Areebi Index Q2 2026 for the global governance-maturity picture.

The human-error and runtime-control case

Approximately 59% of H1 2025 notifiable breaches came from malicious or criminal attack and around 37% from human error¹. The human-error share is the single most important number in this report for AI data governance, because human error is the failure mode that AI adoption multiplies and that runtime controls are built to intercept.

Generative-AI tools change the shape of human error. The same well-meaning employee who once emailed a spreadsheet to the wrong recipient can now paste a customer dataset into a public chatbot, or upload a patient record to an unsanctioned summariser. The OAIC data does not isolate AI-channel human error, and we do not claim it does; the point is structural. Around 37% of Australian notifiable breaches already trace to human error¹ before AI tools become ubiquitous in the workflow, and AI adds new, high-bandwidth ways for that error to leak sensitive data.

The cost side reinforces the case. The ASD reported that the average self-reported cost of cybercrime to Australian business rose approximately 50% to about AUD 80,850, with small business at around AUD 56,600, medium business at around AUD 97,200, and large business losses rising approximately 219% to about AUD 202,700². Every human-error breach a runtime control prevents is a cost avoided against a baseline that is climbing fast.

For where runtime enforcement sits in a broader governance programme, see the Areebi platform and the global maturity framing in the Areebi Index Q2 2026.

The responsible-AI execution gap

If breaches describe the problem, the Responsible AI Index describes how far Australian governance has to go to meet it. The National AI Centre and Fifth Quadrant Responsible AI Index 2025 found only approximately 12% of Australian organisations are "leading" on responsible AI - up around four points on the prior edition - with about 17% "emerging"³. The improvement is real, but the level is low: roughly nine in ten organisations are not yet leading.

The depth of the gap shows in practice adoption. The 2024 edition of the Index found that Australian organisations implement on average only about 12 of 38 responsible-AI practices, from a survey of 413 Australian decision-makers³. Implementing under a third of the recognised practices is not a governance programme; it is a partial one. When that partial posture meets the breach landscape of Finding 1 and the human-error exposure of Finding 2, the result is an execution gap: adoption is running ahead of the controls, inventory, and accountability that should accompany it.

The execution gap is the reason a runtime control plane matters more in Australia than a policy library alone. Closing the distance between roughly 12 of 38 practices³ and a defensible posture is tractable with the right platform; see the Areebi platform and the broader Areebi Index Q2 2026 for the maturity model that frames the same gap globally.

What it means under Australian regulation

The breach landscape, the human-error exposure, and the responsible-AI execution gap all land into a tightening Australian regulatory calendar. Two deadlines matter most for regulated organisations governing AI data flows. For the full calendar of Australian AI obligations and commencement dates, see our Australian AI Regulation Tracker.

Privacy Act automated-decision transparency, 10 December 2026

The Privacy Act reforms introduce automated-decision-making (ADM) transparency obligations that commence on 10 December 2026, requiring organisations to disclose in their privacy policies where automated systems - including AI - are used to make decisions that significantly affect individuals. For sectors already carrying the breach exposure documented in Finding 1¹, ADM transparency adds a documentation and inventory burden that a partial responsible-AI posture³ is not yet equipped to meet. For the detail, see our Privacy Act ADM transparency brief.

APRA CPS 230 and CPS 234

For APRA-regulated entities, CPS 230 (Operational Risk Management) and CPS 234 (Information Security) impose obligations on operational resilience and information-security controls that extend to AI data flows and the third parties that handle them. With finance the second-most-breached sector in H1 2025¹ and the average cybercrime cost to business up approximately 50%², the CPS 230 and CPS 234 control expectations and the runtime-control case of Finding 2 point in the same direction. See our APRA CPS 230 for AI brief and the broader compliance library.

The synthesis

Read together, the three datasets make one argument. Australia's most-regulated sectors are both the most-breached¹ and the fastest AI adopters; human error remains a leading breach cause at around 37%¹; the cost of failure is rising sharply²; and only approximately 12% of organisations are leading on responsible AI³. The regulatory deadlines arriving in 2026 do not create this gap; they make closing it non-optional. The most direct lever is a runtime AI data-control layer that intercepts human error and enforces sanctioned tooling, paired with the inventory and accountability the responsible-AI practices describe.

For the global maturity framing of the same execution gap, see the companion Areebi Index Q2 2026, and for the platform built to close it, see the Areebi platform.

References and dataset

The following primary sources underpin the analysis. Each is publicly accessible at the URL listed. We have not modified any number from its original source; every statistic in this report traces to one of the three datasets below.

1. 1. Office of the Australian Information Commissioner, Notifiable Data Breaches Report: January to June 2025. Source for the approximately 532 notifications (down around 10% on the previous half), approximately 59% malicious or criminal attack, approximately 37% human error, and most-affected sectors (health around 18%, finance around 14%, Australian Government around 13%). oaic.gov.au (Notifiable Data Breaches publications) and the statistics dashboard.
2. 2. Australian Signals Directorate, Annual Cyber Threat Report 2024-25. Source for the over 84,700 cybercrime reports (about one every six minutes), over 1,200 cyber security incidents responded to (up approximately 11%), average self-reported cost of cybercrime to business up approximately 50% to about AUD 80,850, large business average loss up approximately 219% to about AUD 202,700, small business about AUD 56,600, and medium business about AUD 97,200. cyber.gov.au (Annual Cyber Threat Report 2024-25)
3. 3. National AI Centre and Fifth Quadrant, Australian Responsible AI Index 2025(with the 2024 edition for the practice-adoption baseline). Source for the approximately 12% of organisations "leading" on responsible AI (up around four points), approximately 17% "emerging", and the 2024 finding that organisations implement on average only about 12 of 38 responsible-AI practices (survey of 413 Australian decision-makers). industry.gov.au (National AI Centre) and Fifth Quadrant report