APRA has published its annual report for 2025-26. Most coverage will focus on the big numbers. Assets under APRA supervision passed $10 trillion for the first time. Superannuation now holds about $3 trillion across 23 million member accounts. [1]
The more useful part is the section on operational risk. It says what APRA found in the first year of CPS 230, what it thinks of how entities are handling AI, and what it expects to see improve next year. If you sit on a board or run risk at an APRA-regulated entity, that is your agenda for the next twelve months. If you are not APRA-regulated, it is still a good guide to where other regulators are heading.
Operational resilience was the priority
APRA calls operational resilience “a critical priority” for the year. It says risk management practice has improved, but cyber threats, reliance on third parties and fast technology change are creating new weak points. Its words are plain: “As risks evolve, so must industry practices.” [1]
CPS 230 Operational Risk Management started on 1 July 2025. It was a key supervisory focus all year. APRA ran targeted reviews of large entities to test how they had put the standard in place. Where it found material risks or shortcomings, it issued requirements to fix them “in a timely manner.” [1]
That is the first lesson. APRA did not wait to see how CPS 230 settled in. It tested it in year one, and it wrote formal remediation requirements where entities fell short.
APRA now has a map of who relies on whom
In October 2025, every regulated entity had to give APRA, for the first time, a complete list of the material service providers it relies on for critical operations. APRA then worked with entities on how they manage those relationships, with a focus on cases where the dependency was rated high. [1]
Put that next to another piece of work. On 30 June 2026 APRA published the results of its first System Risk Stress Test. The test included a disruption at a material service provider on top of liquidity and market shocks. The banks and super funds that took part came through it. But APRA named “common dependencies on major service providers” as one of the things that could make a crisis worse. [1][2]
So APRA now holds two things it did not have two years ago. It has a register of material service providers across the whole regulated system. And it has its own evidence that shared providers are a system risk. It can see where many entities depend on the same few providers.
The next question follows from that. Your list told APRA who you rely on. It did not show how well you manage each of them. Expect APRA to ask for that: the risk assessment behind each material provider, how often it is reviewed, what the contract allows, and what you would do if that provider failed tomorrow.
Your service provider list told APRA who you rely on. It did not show how well you manage them. That is the next question.
43 incident notifications, and why the small ones matter
CPS 230 requires entities to tell APRA about operational risk incidents and disruptions. APRA received 43 notifications during the year. It assessed most of them as low risk. It still followed them up, because, in its words, they “have the potential to create more material problems if left unchecked.” [1]
This is worth noting. APRA is not only interested in the serious incidents. It reads the low-level ones as a sign of how well controls are working. A run of small notifications in the same area tells APRA that something underneath is not right. Your incident register is being read as evidence of control health, not just as a log.
AI: APRA has said what it expects, and when
On 30 April 2026 APRA wrote to all regulated entities about artificial intelligence. The annual report repeats the core message. Governance, risk management and operational resilience practice are not keeping pace with the scale, speed and complexity of AI adoption. [1][3]
The letter was specific about the gaps. Some entities rely on a single provider for several AI uses, with weak contingency plans. AI is increasingly built into vendor software and developer tools, which makes it harder to see how a model is trained, updated or limited. Information security practice is struggling to keep up. APRA also warned that frontier AI models could make it easier for attackers to find vulnerabilities, which raises the likelihood, speed and scale of cyber attacks. [3]
Two lines matter most for boards. APRA said many boards “lack the technical literacy required to provide effective challenge to management” on AI risk. And it said it expects entities to “demonstrate clear improvements in the next reporting period.” [1][3]
APRA has not created a new AI standard. It expects AI risk to be managed under the standards that already exist, covering information security, operational risk, governance and data risk. [3] That means there is no new rulebook to wait for. The test will be applied using the obligations you already have, and the clock has already started.
There is no new AI rulebook to wait for. APRA will test AI risk against the standards you already have, and it expects to see improvement this year.
Cyber stayed on the list
Cyber resilience runs through the whole report. For superannuation trustees, APRA reinforced its expectations on authentication, information security controls and the uplift needed to protect members. [1]
APRA and ASIC also ran joint industry roundtables on emerging cyber and frontier AI risks. APRA describes this as a single coordinated forum in place of separate engagement processes. [1] For entities that answer to both regulators, it is a sign that the two are talking to each other and comparing what they hear.
Non-financial risk has a capital price
The enforcement section shows what happens when non-financial risk management falls short. During the year APRA applied a $50 million operational risk capital add-on to one bank for weaknesses in non-financial risk management and risk culture. It applied a $2 million capital requirement to an insurer for weaknesses in non-financial risk and regulatory reporting. It also took action against ten entities and three individuals across governance, accountability, expenditure and risk management. [1]
The same section shows the reverse. APRA removed a $500 million operational risk capital add-on from a major bank, first applied in 2019. It came off after a multi-year risk transformation that addressed APRA’s concerns about culture, governance and accountability. [1]
The message is clear in both directions. Weak operational risk management now costs capital. Fixing it properly, and proving it, earns the capital back. It also takes years, not months.
What to do now
Start with your material service providers. You have already told APRA who they are. Now make sure you can show, for each one, a current risk assessment, the evidence behind it, and a tested plan if it fails. Pay most attention to the providers you share with much of your industry, because APRA has said that is where system risk sits.
Next, find the AI you did not choose. Most organisations know about the AI projects they started. Fewer know where AI is built into the software and services they already buy. APRA has named that as a gap. An inventory of AI across your third parties is the obvious first step.
Then look at what your board can actually challenge. APRA has said directly that boards need more technical literacy on AI. That is not about turning directors into engineers. It is about giving them reporting in plain business terms, so they can ask the right questions and judge the answers.
Finally, treat your incident register as evidence. If APRA reads small incidents as signs of control weakness, so should you. Look for patterns before APRA does.
Where BayRisk lands on this
None of this is a project you finish. A service provider register goes out of date the day a contract changes. AI turns up inside a product through a routine update. Incidents keep coming. APRA is asking for a risk operation that runs all year and can show its position at any time, not a compliance exercise done once before a review.
That is what we mean by RiskOps. BayRisk provides the senior risk layer that keeps that operation running. We track third-party risk on evidence, tie it to the business outcomes it puts at risk, and report it to boards in language they can challenge. For many organisations that is cheaper and faster than building the whole capability in house.
If you want to know how your service provider and AI risk position would stand up to APRA’s next round of questions, that is a short conversation: hello@bayrisk.com.au.
References
- Australian Prudential Regulation Authority, Annual Report 2025-26, 2026. https://www.apra.gov.au/news-and-publications/apra-annual-reports
- Australian Prudential Regulation Authority, “APRA publishes findings of inaugural System Risk Stress Test”, media release, 30 June 2026. https://www.apra.gov.au/news-and-publications/apra-publishes-findings-inaugural-system-risk-stress-test
- Australian Prudential Regulation Authority, “APRA calls for a step change in AI-related risk management and governance”, media release and letter to industry, 30 April 2026. https://www.apra.gov.au/news-and-publications/apra-calls-step-change-ai-related-risk-management-and-governance