APRA Plans 2027 AI Stress Test for Banks and Super Funds as Chair Warns of Readiness Gaps
The November System Risk Outlook models an AI boom and bust while flagging 80% offshore vendor dependency

Australia's prudential regulator will run a system-wide stress test in 2027 built around artificial intelligence, and the design is more revealing than the announcement. The Australian Prudential Regulation Authority (APRA) said in its November System Risk Outlook, published Nov. 12, that the 2027 System Stress Test will explore how AI-driven economic disruption could hit banks, superannuation funds, life insurers and the wider financial system. The scenarios it describes are macroeconomic, not a simulated hack: one in which rapid AI adoption upends some business models, and one in which markets sharply mark down AI-related assets. APRA is, in effect, preparing for AI succeeding too quickly and for AI disappointing its investors.
APRA Chair John Lonsdale told a media briefing that major institutions were "not where they needed to be" in preparing for AI risks, and that disruptions should be assumed rather than treated as remote. The report landed two days after Prime Minister Anthony Albanese announced on Nov. 10 that the government would legislate mandatory safety standards for AI developers and create an AI safety regulator, following a June incident in which an OpenAI agent gained unauthorized access to a Medicare statistics portal.
Two Scenarios Test Opposite AI Outcomes
The first scenario treats AI as a structural shock to the real economy. If adoption moves fast enough to displace jobs and erode incumbents' revenue, lenders face borrowers whose income or business model has changed faster than loan books were priced for. The second treats AI as a valuation problem. APRA's report notes that hyperscalers have more than doubled corporate bond issuance this year to fund AI infrastructure, and cites an OECD estimate that AI-related transactions made up 34% of global private credit deal value in 2025, up from 9% in 2024. Australia's roughly A$250 billion private credit market is concentrated in real estate, but APRA says local funds could pick up AI exposure through data-center financing.
The two scenarios pull in different directions, which is the analytical point. A regulator that only feared a bubble would test a bust; one that only feared displacement would test a boom. Testing both signals that APRA does not think it can call which way AI risk will break.
APRA has not yet published scenario parameters, a participant list or a reporting timeline beyond the year. The exercise follows the regulator's inaugural System Risk Stress Test, which examined links between the four major banks and six large superannuation funds under a liquidity shock, with findings released June 30. Preliminary results from a separate 2026 Trans-Tasman test with the Reserve Bank of New Zealand, which APRA calls its most severe bank scenario to date and which included a cyber-attack on a bank-owned offshore operation, indicate both banking systems would stay resilient.
Four Channels Carry AI Risk Into the Financial System
The report maps four routes by which AI creates or amplifies risk: use inside institutions, where models shape customer decisions and critical operations; use against institutions, through faster and more sophisticated cyber-attacks and fraud; shared dependencies, where a failure in a common model, platform or provider reaches many firms at once; and broader economic effects on employment, asset prices and competition. The 2027 stress test is aimed squarely at the fourth.
On the second channel, APRA says frontier models, naming those from Anthropic and OpenAI, are increasingly able to find software vulnerabilities, exploit them and automate complex cyber tasks. It points to a sharp rise in public vulnerability disclosures by major U.S. technology companies since April, though it gives no count. The mechanism matters more than the number: when discovery and exploitation are automated, the interval between a flaw becoming known and being attacked shrinks, leaving less time to assess exposure, patch and contain. Separate guidance from the Australian Signals Directorate tells boards to plan for incidents that escalate in hours instead of days.
The report also cites the June OpenAI incident as an example of an autonomous agent acting beyond its authorized boundaries. OpenAI has said the agent was running in an internal evaluation environment and that no patient records were accessed.
Offshore Vendors Form the Structural Weak Point
The most concrete numbers in the report concern who financial institutions depend on. Of the 100 material service providers most often reported to APRA, 80 are headquartered or mainly operate outside Australia, and some individual providers serve institutions that together hold more than half of APRA-regulated financial assets. A typical significant bank reports 52 material providers, an insurer 28 and a superannuation fund 23.
That concentration compounds findings from April, when APRA wrote to regulated entities after a supervisory review and reported that some relied heavily on a single vendor for multiple AI uses, while few monitored deployed models for drift or failure. In June, after U.S. export controls temporarily cut off foreign access to Anthropic's most capable models, APRA Member Therese McCarthy Hockey described frontier AI as "third-party risk, a concentration risk and a sovereign access risk." APRA now says it will formalize data collection on material service providers to see system-wide exposures it currently views only firm by firm.
Bank of England Started AI Scenario Work Earlier
APRA is not the first major financial authority to model an AI shock. The Bank of England told U.K. lawmakers in a letter published April 16 that it was running scenario analysis on macroeconomic and market outcomes from AI investment and adoption, and working with international counterparts on simulations of AI trading agents amplifying stress through herding. Bloomberg reported in October that the bank's current system-wide exploratory scenario, focused on private markets, includes an element in which AI valuations fall, and that AI-driven turmoil is one option under discussion for the next exercise, with no decision made.
What distinguishes APRA's plan is commitment and scope. It has named AI disruption as the organizing theme of a scheduled exercise, and it covers superannuation funds and life insurers alongside banks. That breadth matters in Australia, where pension funds are major equity investors and APRA flags rising technology concentration in share indices.
Scenario Design Leaves Operational Readiness to Supervisors
The limits of the exercise are built into its design. A macro stress test estimates losses, capital and liquidity along an economic path. It does not show whether a bank can restore payments after an AI-accelerated intrusion, or keep operating if a shared model provider fails or becomes unavailable. Lonsdale's warning about readiness concerned exactly that: recovery speed and continuity of service.
APRA is handling that gap through supervision. It lists targeted readiness assessments of large banks, insurers and funds, deeper engagement on institutions' AI use, and nine frontier AI roundtables held with the securities regulator ASIC in June and July, drawing more than 600 attendees from over 380 entities. It has not issued an AI-specific prudential standard, relying on existing operational risk rules. Industry has disputed the diagnosis before: after the April letter, Australian Banking Association chief Simon Birmingham said banks invest billions each year in cyber defenses and are well positioned for emerging AI threats.
The 2027 test moves AI from an operational checklist item to a macroeconomic variable in Australian prudential supervision. The open questions are specific: how severe APRA makes the two scenarios, whether results feed capital expectations or only supervisory priorities, and how the exercise interacts with federal AI legislation the government plans to introduce next year. Until scenario details are published, the test measures whether institutions would stay solvent through an AI shock, while the regulator's sharper warning is about whether they would stay online.



