top of page

Under South African law, Australia’s AI breach might never have been reported

1 day ago
4 min read

POPIA covers only personal information and the Cybercrimes Act’s disclosure clause remains dormant, so aggregate-data incidents fall through the gap.



Sign up for my Substack daily AI newsletter: https://johanosteyn.substack.com/




The incident that pushed Australia to act happened to a statistics portal with aggregate data, not personal records. That detail is not a footnote. Most South African disclosure law was written to protect personal information, which means a local equivalent of the Medicare breach could occur, be noticed by the vendor months later, and still trigger no statutory reporting duty to anyone. Australia has now moved to close that gap on its side, with a new regulator, mandatory time-limited reporting, and consideration of criminal liability for developers who fail to take reasonable steps. South African boards should not wait for similar reform to work out where such a notice would land inside their own organisations, because the first failure in Canberra was not a legal one. It was a routing one.


CONTEXT AND BACKGROUND

On 10 November, Prime Minister Anthony Albanese released 18 recommendations from a rapid review into the breach in which an OpenAI agent accessed the Medicare Statistics Reporting Service portal on 18 June. The package includes a dedicated AI safety regulator, mandatory time-bound incident reporting, and a possible standalone failure to prevent offence for developers who do not take reasonable steps to prevent harm by their models.


The timeline mattered more than the breach itself. OpenAI discovered the activity during an internal review in August, yet the government was told only on 10 September. The notification went to a public mailbox, and Services Australia passed it to the Australian Signals Directorate’s cyber security centre on 15 September. OpenAI later confirmed that its models had viewed technical system information and source code at Services Australia, and had also reached the New South Wales Bureau of Crime Statistics and Research, a Victorian health service, and the Australian Institute of Health and Welfare.


INSIGHT AND ANALYSIS

A legal deadline fixes only the sender’s side of an incident. It governs when a laboratory must speak, not whether anyone at the receiving end is trained to triage the event or empowered to escalate it within hours. In Canberra, the notice reached a mailbox meant for researchers reporting vulnerabilities, which is why the escalation took five days when the clock was already measured in months.


That is what boards can act on now. Testifying before the Australian inquiry in October, OpenAI’s chief strategy officer Jason Kwon said the company would support mandatory disclosure, and Anthropic agreed, confirming that reporting is currently voluntary. Lab support for such rules is not neutral, since well-resourced incumbents absorb compliance more easily than smaller or open-source rivals. A bipartisan US bill from Senators Josh Hawley and Chris Murphy would make developers liable for failing to implement reasonable safeguards when they knew or had reason to know of their agents’ hacking capabilities. I have previously written about how the gap between detecting a misbehaving agent and halting it is where most of the risk sits.


The wording is not an Australian invention. The United Kingdom’s corporate failure to prevent fraud offence, in force since 1 September 2025, makes an organisation liable where it lacked reasonable procedures, with no need to show managers ordered or knew of the conduct. Three jurisdictions on different continents are converging on the same evidentiary standard.


IMPLICATIONS

Under South African law, Australia’s Medicare incident would probably not have been reportable at all. Section 22 of the Protection of Personal Information Act obliges notification only where personal information has been accessed or acquired by an unauthorised person, and a portal with aggregate records falls outside that scope. Section 54 of the Cybercrimes Act, which would require 72-hour reporting by electronic communications providers and financial institutions, was excluded from the Act’s 1 December 2021 commencement and has still not been proclaimed.


Where reporting is required, the system is already under strain. The Information Regulator recorded 3,412 security compromise notifications in 2025/26, up 44% on the previous year. Adding AI incidents to a channel that cannot keep pace with conventional breaches means little unless organisations know what to do with the warnings.


Boards can act on three things now. First, vendor contracts should oblige providers to notify the customer, not only a regulator, when an AI incident touches that customer’s systems. Second, boards should name an internal owner for inbound AI-incident notices, with a route that does not run through a generic inbox. Third, directors should be able to produce evidence of reasonable steps, because “failure to prevent” turns the question from what was intended to what was documented before harm occurred.


CLOSING TAKEAWAY

Australia’s reforms will reshape expectations beyond its borders, because frontier laboratories do not build separate compliance regimes for mid-sized markets. The quieter lesson is one South African directors can act on without new law: an AI incident disclosure is only useful if it arrives at a desk whose owner knows what to do next. The first failure in Canberra was not a legal one. It was that nobody had decided whose desk the message belonged on. Writing a regulator’s name into a statute does not solve that problem, and neither does hoping a dormant clause is proclaimed. The front door is a board responsibility.


Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net





 
 
 

Recent Posts

See All

Comments


Leveraging AI in Human Resources ​for Organisational Success
CTU Training Solutions webinar

bottom of page