Key takeaways
- Australia is moving towards mandatory AI incident reporting. Following an AI agent’s unauthorised access to government websites, the Australian Government has accepted 18 recommendations covering incident notification, accountability, cyber security and resilience.
- AI incidents create new challenges for detection, disclosure and investigation. Organisations need to understand what autonomous systems can access, recognise unexpected behaviour and ensure incidents are reported promptly to the appropriate parties.
- New Zealand professional services firms should review their AI supplier arrangements and incident response plans. Clear reporting thresholds, contractual notification obligations and reliable investigation records are becoming increasingly important.
On 10 November 2026, the Australian Government announced its response to an incident involving an OpenAI agent that accessed Australian Government websites, including the Medicare Statistics Reporting Service Portal. Following a rapid review commissioned in September, the Government accepted 18 recommendations intended to strengthen Australia’s ability to prevent, detect and respond to AI-driven cyber incidents. The recommendations include mandatory reporting by AI companies of certain incidents involving unexpected or misaligned behaviour, stronger resilience and response arrangements, and consideration of criminal-law reform for serious misconduct.
For New Zealand professional services firms, the development raises an important governance question: if an AI system behaves unexpectedly or accesses information outside its intended scope, how quickly would the organisation know, and who would be responsible for reporting it?
What happened in Australia?
The Australian Government commissioned its review after OpenAI notified authorities that an AI agent had accessed government websites, including a health-related statistics portal. The incident attracted particular concern because it involved autonomous AI activity affecting government systems and questions about the adequacy and timing of notification. The Government’s announcement does not establish that patient medical records were accessed. Public reporting indicates that the review found no evidence of patient information being obtained, although the full technical circumstances have not been publicly disclosed.
This distinction matters. An unauthorised access event does not automatically establish that sensitive information was stolen, but it may still reveal weaknesses in system boundaries, monitoring or incident reporting. The incident demonstrates that organisations need processes for identifying and investigating unexpected AI behaviour even where the initial evidence does not establish data loss or malicious intent.
Australia proposes stronger AI incident reporting
The Australian Government has accepted all 18 recommendations from the rapid review and commenced implementation work. A central recommendation is mandatory reporting by AI companies when certain misaligned behaviour or cyber incidents occur. This would move reporting beyond arrangements that depend largely on providers deciding when and how to disclose an event. The Government is also considering criminal-law reform for serious misconduct, stronger engagement with international AI developers and improvements to cyber incident response and information sharing.
These proposals represent a significant policy direction, but they should not be confused with legislation already in force. The announcement does not establish a general mandatory AI incident reporting regime for New Zealand businesses, nor does it mean every unexpected AI output would become a reportable event. The practical importance lies in the expectation that organisations developing and operating increasingly autonomous AI systems should be able to identify serious incidents, disclose them appropriately and support subsequent investigation.
Why AI incidents can be difficult to recognise
Traditional cyber incidents often involve identifiable indicators such as compromised credentials, malicious software, unauthorised access or suspicious network activity. AI agents can complicate this picture because they may operate through legitimate accounts, approved applications and authorised connections while taking actions that exceed their intended purpose. An agent instructed to research information might follow unexpected links, retrieve restricted material or interact with systems outside its intended workflow. Similarly, an automated document-processing agent might access the wrong client matter or transmit information to an unintended destination.
These situations do not necessarily involve a conventional external attacker. They may arise from unexpected model behaviour, inadequate permission boundaries, poorly designed workflows or weaknesses in connected systems. For incident response teams, the investigation may therefore require more than identifying whether credentials were compromised. It may also require reconstructing the agent’s instructions, tool use, decisions and sequence of actions.
The importance of timely supplier notification
Most New Zealand professional services firms do not develop their own frontier AI models. They rely on commercial AI platforms, software providers and increasingly sophisticated integrations. This creates a dependency on suppliers to identify incidents and communicate relevant information promptly. If an AI provider discovers that its agent has accessed information unexpectedly, affected customers may need to assess whether confidential information was exposed, whether connected systems remain secure and whether regulatory or contractual notification obligations apply.
Delayed disclosure can make those assessments more difficult, particularly where logs have limited retention periods or potentially affected information cannot easily be reconstructed. Organisations should therefore examine what their AI supplier agreements say about security incidents, unexpected agent behaviour, notification timeframes, investigation cooperation and access to relevant evidence. A general assurance that a provider maintains appropriate security controls is not a substitute for understanding what happens when those controls fail.
What this means under New Zealand law
Australia’s announcement does not create new legal obligations for New Zealand organisations. However, existing New Zealand privacy, security and professional responsibilities remain relevant when AI systems cause incidents. Under the Privacy Act 2020, organisations must notify the Privacy Commissioner and affected individuals as soon as practicable after becoming aware of a notifiable privacy breach. The notification threshold concerns breaches that have caused, or are likely to cause, serious harm.
An unexpected AI action does not automatically meet that threshold. The organisation must assess the circumstances, information involved, potential consequences and applicable legal requirements. Professional services firms may also have confidentiality duties, client contractual obligations, cyber insurance notification conditions or sector-specific reporting requirements.
For example, a law firm using AI to review client documents would need to investigate any unexpected disclosure of confidential material, even where the event did not meet the statutory threshold for a notifiable privacy breach.
The central lesson is that AI incidents should be assessed through established legal and professional frameworks while recognising the additional evidence and technical questions created by autonomous systems.
Incident response needs to include AI behaviour
Many organisations already maintain cyber incident response plans covering ransomware, account compromise, data breaches and service disruption. Those plans provide a useful foundation, but they may not adequately address incidents involving AI agents. An effective investigation may need to establish the agent’s assigned task, available permissions, instructions, connected applications, information accessed and actions completed.
It may also be necessary to determine whether the agent can be stopped, whether other automated processes remain active and whether affected information can be recovered or corrected. Where an external provider operates the system, the organisation may depend on that provider for technical records and containment assistance. These considerations should be addressed before AI agents are given access to sensitive client information or business-critical applications.
What this means for your organisation
1. Identify AI systems capable of taking action. Maintain an inventory of AI agents, connected applications and automated workflows. Record their purposes, permissions, suppliers and accountable business owners.
2. Define AI incident reporting thresholds. Establish which events require internal escalation, including unexpected access, unauthorised changes, inappropriate disclosure and actions outside approved boundaries. Distinguish technical anomalies from events requiring legal or regulatory assessment.
3. Review supplier notification arrangements. Confirm how providers will notify the organisation of relevant incidents, what information they will provide and whether contractual arrangements support timely investigation.
4. Preserve investigation evidence. Determine whether instructions, tool calls, access logs, system changes and relevant approvals can be reconstructed. Apply appropriate retention and access controls to protect confidential information.
5. Test containment procedures. Establish how an AI agent can be suspended, disconnected or have its permissions revoked. Test these controls under realistic conditions rather than assuming that disabling an application will stop every connected process.
6. Update incident response exercises. Include scenarios involving unexpected AI access, confidential information exposure and delayed supplier notification. Test technical containment, legal assessment, client communications and management decision-making.
7. Establish clear accountability. Assign responsibility for investigating AI incidents and deciding when external notification is required. Ensure legal, privacy, security and operational teams can coordinate effectively.
A practical direction for AI governance
Australia’s response illustrates a developing expectation that AI providers and operators should be accountable not only for preventing incidents, but also for recognising, reporting and investigating them. For New Zealand professional services firms, this is a useful opportunity to review existing arrangements before autonomous AI becomes more deeply integrated into client services. The priority is not to create a separate reporting process for every AI error. It is to ensure that significant unexpected behaviour can be identified, contained, investigated and assessed against existing obligations. As organisations delegate more tasks to AI, the ability to explain what happened and respond appropriately when something goes wrong will become an increasingly important measure of effective governance.
Sources
- Australian Prime Minister, Strengthening Australia against AI-driven cyber incidents, 10 November 2026
- Reuters, Australia to mandate AI breach reporting and create regulator after OpenAI breach, 10 November 2026
- New Zealand Privacy Act 2020, official legislation

