APRA puts AI at the centre of financial risk: what NZ firms check

By DataForge team18 November 20265 min read

APRA's November 2026 risk outlook names AI as a central threat to financial stability. NZ firms can borrow its four risk channels to test their own controls.

What APRA published on 12 November

The Australian Prudential Regulation Authority released its System Risk Outlook for November 2026 on 12 November. The twice-yearly report covers the banks, insurers and superannuation funds APRA supervises. This edition makes AI the main theme. At a media briefing the same day, APRA chair John Lonsdale said it's "pretty safe to assume there will be interruptions, there will be problems", according to ABC News. He also said major institutions weren't yet adequately prepared for AI risk.

The report sorts AI risk into four channels. Institutions use AI in their own decisions and operations. Attackers use AI against institutions to speed up attacks and fraud. Many firms rely on the same models, platforms and providers, so one failure can hit several of them at once. And AI could have wider effects on jobs, asset prices and business models. APRA's prudential framework is technology neutral, and the regulator says it doesn't regulate AI technology itself. APRA supervises AI through its existing governance, information security, operational resilience and third-party risk requirements, and it wants institutions to show evidence that they've identified and are managing emerging AI risks.

The cyber section is the most immediately relevant. APRA says frontier models from Anthropic and OpenAI are speeding up the discovery and exploitation of software vulnerabilities. It also says major US technology companies have reported a sharp rise in public vulnerability disclosures since April 2026. Attackers can therefore move sooner, and institutions get less time to assess exposure and patch. The report also cites the June 2026 case where an OpenAI agent got around controls to reach an Australian Government website. APRA warns that agents without human oversight and preventative controls can cause harm their operators never intended.

Findings from the roundtables and supplier data

APRA and ASIC ran nine roundtables between June and July 2026 with more than 380 entities and 600 attendees, as reported by Insurance Business. The regulators told attendees that awareness and written plans fall short of what's needed. Institutions need governance, controls and response capability that hold up under pressure. The five priorities that came out of those sessions start with cyber basics such as timely patching, identity and access controls, and tested recovery. They also cover governance and escalation that can act on incomplete information, careful use of AI in defence with human oversight, mapping shared dependencies, and closer collaboration on threat intelligence and supplier assurance.

The supplier data is a useful reference point for anyone thinking about concentration risk. Under Prudential Standard CPS 230, regulated institutions report their material service providers. Of the top 100 providers reported, 80 are headquartered or mainly operate outside Australia. A typical significant bank reported 52 material providers, a typical insurer 28 and a typical super fund 23. APRA says some providers support institutions holding more than half of APRA-regulated financial assets.

APRA's 2027 System Stress Test will look at AI-driven economic disruption for the first time. The outlook sets out downside cases where fast AI uptake upends some business models, or where markets sharply mark down the value of AI-related assets. APRA hasn't published the scenario numbers.

Why the report matters on this side of the Tasman

The New Zealand connection is direct. APRA and the Reserve Bank of New Zealand recently completed the 2026 Trans-Tasman Stress Test, the first joint exercise since 2017. The test covered five large Australian banks and four large New Zealand banks, and APRA says the two regulators now share information more closely. APRA plans to publish more detailed results later this year.

New Zealand's regulators had already taken a similar position. On 15 October 2026 the RBNZ and the Financial Markets Authority set out supervisory expectations on cyber risk from frontier AI. Those expectations also build on existing prudential, conduct, cyber and operational resilience obligations, and they start with getting foundational cyber controls right. The FMA's Louise Unger said boards and senior management need to know where AI is used in their business and who's accountable for its risks. The two regulators use different wording, but they agree on the same points: existing rules already cover AI, boards own the risk, and cyber basics come before anything new.

Most NZ businesses aren't supervised by APRA, the RBNZ or the FMA. Two practical consequences still reach them. Firms that supply banks, insurers or KiwiSaver providers can expect tougher supplier assurance questions, because both regulators have named third-party and concentration risk as a focus. And APRA's four channels work as a ready-made checklist for any organisation using AI. A distributor running an AI agent against its ERP faces the same failure modes as a bank, at a smaller scale. That includes an agent acting beyond its brief, an attacker exploiting an unpatched system faster than before, and a single model provider outage stopping several business processes at once.

APRA hasn't said what evidence will satisfy supervisors that AI risks are "effectively identified and managed", and New Zealand's regulators haven't published a checklist either. Firms that build a defensible record now will be better placed when the questions arrive.

Next steps for NZ owners and IT leaders

  • Build an AI use register. List every AI tool and agent in use, including features switched on inside existing software. Record the business owner and the person accountable for each one's risks, in line with the FMA's board-level expectation. Our AI consultancy team uses a register like this as the starting point for governance work.
  • Map shared dependencies. Note which model providers, cloud platforms and SaaS vendors sit under each AI use. Work out what stops if one provider is unavailable for a day, and where those providers are based.
  • Shorten patch windows. APRA expects attackers to exploit published flaws faster. Compare your patch targets for internet-facing systems with that expectation, and treat critical vulnerabilities as change requests with a deadline. A structured ITIL 4 change and incident process makes those deadlines enforceable.
  • Put preventative controls on agents. Any agent that can write to external systems or spend money needs scoped credentials, an action allowlist and human approval for consequential steps. Review agent logs on a set schedule.
  • Rehearse escalation with incomplete information. Run a tabletop exercise where an AI-related incident is unclear for the first few hours. Check who can decide to pull a tool offline and how quickly they can be reached.
  • Prepare supplier evidence. If you serve financial institutions, assemble your AI register, access controls and incident response plan before their assurance questionnaires arrive.

Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.

Get my free assessment How we help
Get new articles by RSS