The Australian Prudential Regulation Authority (APRA) says the use of artificial intelligence (AI) against the institutions it regulates has worsened over the past year and that governance inside those institutions has still not caught up with the technology they are deploying.
APRA Chair John Lonsdale (pictured) gave the assessment at a media briefing today, held after the regulator published the third edition of its System Risk Outlook.
Asked how insurers, banks and superannuation trustees had responded to the letter APRA issued to industry on artificial intelligence risk in April 2026, Lonsdale was direct.
“There's definitely more work required,” he said. “That letter was really about how entities are deploying it themselves. And there is certainly more work to do to uplift governance and controls. And we've got an active supervision program on that.”
Lonsdale set out three AI focus areas for APRA: whether regulated entities have the governance and controls they need, the use of AI as a weapon against those entities, and dependence on third-party providers. Two of the three have deteriorated since April.
Read next: APRA pushes insurers to narrow AI risk oversight gap
“These other two elements are very important and have escalated, to be blunt,” he said. “So the threat environment and how AI can be used against our regulated entities, that has got worse over the course of the year. And the third-party risk is very much work in progress for us.”
He said the pace of attack was the core concern. “The speed with which AI can be used to identify issues and vulnerabilities and cause problems is something that we're very concerned about,” Lonsdale said, adding that entities were committing “very large amounts of resources” to plugging holes in legacy and operational systems.
The clearest shift in APRA's position was on what resilience means. Lonsdale said the regulator no longer frames it around prevention.
“When we're talking about resilience, we're not just talking about the prevention of a problem, but I think it is pretty safe to assume there will be interruptions, there will be problems,” he said. “And so we're talking to entities about, have you got the systems in place to make sure that you can recover at speed and make sure that there's continuity of service provision to customers?”
On third parties, Lonsdale said APRA's pressure had already changed commercial decisions.
“In some cases, some entities have said, well, I don't want to take that risk, I might insource that,” he said. “So there has been change there.”
Read next: Data centre insurance capacity stretched as regulator warns on AI
The System Risk Outlook reports that of the top 100 material service providers named by APRA-regulated institutions, 80 are headquartered or have their principal operations outside Australia. The typical significant insurer identified 28 such providers. Lonsdale said APRA had met hyperscalers and other providers directly, and that the gap between APRA's visibility and its authority over suppliers remains: “We are not the regulator for a lot of these third parties.”
Asked whether Australian companies should favour United States or Chinese AI systems, Lonsdale said APRA would not take a position.
“We don't have a house view on particular types of models,” he said. “It's not for APRA to say don't use that model or use this model instead, but we certainly need them to understand what these models are capable of doing, what their strengths are, what their weaknesses are, and where the vulnerabilities are.”
He said capability varies sharply by size. “You've got some very large players who are better equipped, to be frank, to do that. And if you're small, it's much harder.”
Lonsdale said the technology landscape had changed materially since April, citing new frontier model releases and the arrival of open-source, open-weight Chinese models. APRA's three concerns, he said, had not changed with it.
Pressed on the June 2026 incident in which an OpenAI agent circumvented controls to access an Australian Government website, Lonsdale declined to assess the company's handling of it but did not play down the event.
“I don't want to comment on specific entities, particularly the ones that I don't regulate, but I think it's fair to say that that's a very significant example of what can go wrong,” he said. “And it's just one of a number, actually, that we're looking at very closely.”
