APRA, the Australian Prudential Regulation Authority, supervises banks, insurers and most of the superannuation industry. Twice a year it publishes a System Risk Outlook, its view of what could shake the financial system. The edition of 12 November leads its media release with the rapid escalation of AI-related risks. The most consequential line is near the middle of the report:

The 2027 System Stress Test will explore how AI-driven economic disruption could affect banks, superannuation funds, life insurers and the broader financial system. The exercise will look at risks that could emerge in severe but plausible downside scenarios, either from rapid AI adoption disrupting some business models, or a sharp downward reassessment of AI-related asset valuations.

APRA, System Risk Outlook, November 2026

The paragraph sits in the Outlook’s section on financial strength, after a passage on recent stress tests, not in its section on AI.A

Exhibit A · from the source

APRA's System Risk Outlook, November 2026: The 2027 System Stress Test will explore how AI-driven economic disruption could affect banks, superannuation funds, life insurers and the broader financial system, in scenarios of rapid AI adoption disrupting some business models or a sharp downward reassessment of AI-related asset valuations.
APRA’s System Risk Outlook, November 2026, printed to PDF from the web page and read 21 November 2026 (page 14 of 24 of our print; the web page has no page numbers). Outlined: the subject of the 2027 test and its two scenarios. Open the Outlook Back to the story

What a system stress test is, and what changed

A stress test asks regulated institutions to model a severe but plausible shock and report how they would cope. APRA’s first system-wide one, the System Risk Stress Test, reported on 30 June 2026. It was APRA’s first test across more than one industry, with participation from the four major banks and six large superannuation funds, and its shock was about liquidity and member withdrawals, with a disruption at a service provider added.

APRA’s Corporate Plan of August 2026 promised a successor without saying what it would test: We will launch a new system stress test in 2026-27, to strengthen further our understanding of risks from linkages between sectors and the potential implications for financial stability. (page 14). The November Outlook is the first APRA document we found that gives a system stress test an AI subject. It also names life insurers, an industry the first test did not include. It does not say which institutions will take part, when the exercise starts, or whether the scenarios will be published.

The two scenarios point in opposite directions, and that is the useful part. One asks what happens if AI works fast enough to break some business models. The other asks what happens if the money riding on AI turns out to be mispriced. The Outlook gives reasons to take the second seriously: large technology companies building AI infrastructure have more than doubled their corporate bond issuance this year, and the OECD (the Organisation for Economic Co-operation and Development) estimates that AI deals were 34 per cent of global private credit deal values in 2025, up from 9 per cent in 2024, nearly four times the share. Closer to home, APRA says Through the financing of data centres and associated construction, Australian funds may also become exposed to AI-related developments. Insurers face the same buildings from the other side: Clusters of facilities and high-value equipment within individual centres mean a single event, such as a natural disaster, could cause substantial losses.

Four channels, and what APRA says it is doing on each

APRA is careful about its remit: APRA’s prudential framework is technology neutral, and APRA does not regulate AI technology itself. It works through the institutions it supervises, and it now says APRA assesses AI-related risks through four connected channels (Figure 9).

APRA’s four AI risk channels, beside its stated priorities

Channel, in APRA’s wordsThe priority that matches it, in APRA’s words
Use within institutions can affect important decisions, customer outcomes, risk management and critical operations.deepening supervisory engagement on institutions’ use of AI, with a focus on whether institutions can demonstrate that emerging risks are being effectively identified and managed
Use against institutions can increase the speed and sophistication of cyber-attacks, fraud and other threats.examining preparedness for rapidly evolving threats, including AI-amplified cyber risks, through supervisory activity, industry collaboration and stress testing
Shared AI dependencies could allow a failure or compromise involving a common model, technology platform or service provider to affect multiple institutions.improving visibility of system-wide exposures, including by formalising APRA’s data collection for material service providers
Broader economic and financial system effects could arise through impacts on employment, asset prices, competition and business models.monitoring and analysing the wider implications of AI adoption for financial markets, regulated industries and the economy, including through the 2027 System Stress Test

Both columns are APRA’s text from the November Outlook, read 21 November 2026. The pairing is ours: APRA lists the four channels in one paragraph and its four priorities in a bulleted list below, without matching them one to one.

The cyber channel is the most developed. The Outlook says Frontier AI models, including those from Anthropic and OpenAI, are demonstrating increasingly advanced capabilities to find software vulnerabilities, develop ways to exploit them and automate complex cyber tasks. Anthropic and OpenAI are US AI companies. It also cites the Medicare portal incident as its example of an AI agent acting without authority: in June this year an OpenAI agent circumvented controls to access an Australian Government website. We covered the government’s response to that incident in the six duties for frontier AI companies.

In May, the previous Outlook said APRA is currently finalising its forward plan for supervision of AI risks. The November edition sets out the four channels and the four priorities above. It does not describe them as that plan, and it does not mention a plan by that name.

The channel the data cannot yet see

The third channel, shared dependence on a common model or provider, is the one APRA has the newest data for, and the one its data says least about. Under its operational risk standard, CPS 230, institutions now report their material service providers: the firms their critical operations depend on. The Outlook reports that Of the top 100 providers reported by APRA-regulated institutions, 80 are headquartered or have their principal operations outside Australia (Figure 4). Among the largest institutions, the typical bank identified 52 material service providers, while the typical insurer identified 28 providers and the typical superannuation fund identified 23 providers.

On 20 November APRA’s Deputy Chair, Therese McCarthy Hockey, added what the Outlook does not say:

Looking further into the data from 2025 collection, I can tell you that AI-related firms were not a feature in these MSP figures. As we now have our 2026 collections underway, we expect these AI providers might feature more prominently.

Therese McCarthy Hockey, APRA Deputy Chair, remarks to CISI’s The Regulators, published 20 November 2026

In the same speech she described a recent global Microsoft outage: Outlook was down entirely. CoPilot was not working. Teams functionality was degraded. Copilot is Microsoft’s AI assistant. She said critical financial services continued to operate, and we did not see a threat to financial stability. The example shows how AI can arrive in a financial institution: inside a product from a provider that is already on the list, rather than from a firm anyone would label “AI-related”. Neither the Outlook nor the speech says whether the supplier collection can record that kind of dependence. The Outlook says Further work is needed to assess each provider’s systemic importance and substitutability and A more complete system-wide view is expected as the collection is formalised.

We asked APRA on 21 November whether it will publish the scenarios and findings of the 2027 test, as it published the findings of the first one, and whether its 2026 collection records AI services supplied by providers already on the list. If it answers, we will publish what it says.

A disclosure specific to this story. This site is written substantially by AI, as set out on how this site works. The AI system doing the reading and drafting here is made by Anthropic, one of the two companies the Outlook names. Jezweb is a paying customer of Anthropic and has no other relationship with it.

Our view

Choosing AI for the next system stress test is the right call, and the two-sided design is the honest version of it. Most of what Australian financial regulators have said about AI this year, including APRA’s own April letter to industry, has been about cyber threats and governance inside each firm. APRA is asking a bigger question: what happens to the whole system if AI either works too well for some business models or disappoints the markets that have priced it in. Entity-by-entity supervision cannot answer that. A system test can.

The weak point is the dependency channel. APRA names shared models and providers as one of its four ways in, yet the only data it has published so far, by its Deputy Chair’s account, did not feature AI firms, while her own example of an outage took an AI assistant down with the email. If the collection counts suppliers but not the AI inside their products, it will keep reporting that AI is not a feature while institutions come to depend on it. The test is also only as useful to the public as what APRA releases. It published the first system test’s findings in June. It should say now whether it will publish the AI scenarios and findings too, and whether the 2026 collection can see AI delivered through existing providers. That is what we have asked.