ASIC Report 798, two years on

What REP 798 found

ASIC reviewed 23 licensees and 624 AI use cases across banking, credit, insurance and advice. Adoption was accelerating and governance was not keeping pace. Nearly half of the licensees had no policies covering consumer fairness or bias. Many relied on their existing risk frameworks without asking whether those frameworks could see an AI model at all.

The report's central position has not moved: the consumer protection provisions, directors' duties and general licensee obligations in the Corporations Act and the ASIC Act already require institutions to govern AI properly. The duty to provide services efficiently, honestly and fairly does not carve out decisions made by a model. No new law is needed for those duties to bite, and ASIC has not published a follow-up thematic report in 2026, which tells you it regards the position as settled.

ASIC in 2026. A new Chair and a cyber letter

Sarah Court commenced as ASIC Chair on 1 June 2026, succeeding Joe Longo. Before the handover, on 8 May 2026, ASIC published an open letter to all AFS licensees and market participants (26-092MR) warning that frontier AI models can expose cyber vulnerabilities at a speed and scale the industry had not planned for. It set out twelve actions, from asset identification and patching to third-party risk and defensive use of AI, plus four expectations of boards, and instructed that the letter be tabled at the board and the risk committee.

The framing is deliberate. ASIC treats cyber resilience as a core licensing obligation, so an AI-accelerated breach that a licensee failed to prepare for is a licence question, not only a technology one.

APRA's letter of 30 April 2026

APRA's letter to industry on artificial intelligence reported the results of a targeted supervisory review of large banks, insurers and superannuation trustees. It did not announce a new prudential standard. It said AI is to be managed under the standards that already exist, chiefly CPS 230 on operational risk, CPS 234 on information security and CPS 220 on risk management, and it named four areas of weakness.

• Security. Prompt injection, data leakage and gaps in managing non-human identities such as agents and service accounts.
• Governance. Not maturing at the pace of adoption.
• Suppliers. Heavy concentration on a few AI providers, with opacity about fourth parties behind them.
• Assurance. Inadequate testing before deployment and weak monitoring for model drift afterwards.

Boards are expected to have enough AI literacy to challenge management and to have set intervention triggers. APRA flagged stronger supervisory action and, where appropriate, enforcement where AI risks are unmanaged. Its May 2026 System Risk Outlook then named AI a system-wide risk.

Resilience at Frontier AI Speed. The joint paper of 27 August 2026

Following their letters, APRA and ASIC convened nine frontier AI roundtables in June and July 2026, attended by more than 600 people from over 380 entities, with the Australian Signals Directorate in support and the Reserve Bank, Treasury and the ACCC in the room. The joint information paper, released with ASIC media release 26-201MR, tells boards and executives to move from awareness to tested action.

Its checklist is written for directors. Get the cyber fundamentals right, meaning asset identification, patching, identity and access controls, backup integrity, tested recovery and third-party risk. Settle risk appetite, escalation authority and crisis communications at board level before an incident rather than during one. Treat concentration in third-party AI providers as a sector-wide exposure, not only your own. Rehearse.

CPS 230 and your AI vendors

CPS 230 is not an AI standard, but since 1 July 2025 it has been the frame AI systems and AI suppliers fall into. A foundation-model provider or AI vendor that supports a critical operation can be a material service provider, which means it must be registered, risk-assessed, covered by contractual terms that meet the standard, and notified to APRA. Transitional relief for contracts that pre-dated the standard ended at the earlier of contract renewal or 1 July 2026, so as at October 2026 every AI supplier that meets the threshold is fully in scope.

The hard date. Automated decision-making disclosure, 10 December 2026

The one AI-adjacent obligation with a legislated start date sits in the Privacy Act. From 10 December 2026, an APP entity that uses personal information in a computer program to make, or substantially contribute to, decisions that could significantly affect a person's rights or interests must say so in its privacy policy, listing the kinds of decisions and the kinds of personal information used. Credit scoring, claims triage, pricing, fraud flagging and onboarding checks are squarely within it.

The OAIC published its final guidance on 30 September 2026: a fact sheet on APP 1.7 to 1.9, a flowchart and an updated Chapter 1 of the APP Guidelines. It reads “computer program” broadly, from rules engines and spreadsheets through to machine learning and generative AI, and confirms that a human signing off does not remove the duty where the program's output is a material input. There is no grace period. The Privacy Act second-tranche exposure draft of 31 August 2026 does not change any of this. Our guide to the disclosure and the free builder cover the drafting.

AML/CTF and AI transaction monitoring

Reformed AML/CTF programs, enterprise-wide risk assessments and updated customer due diligence commenced on 31 March 2026 for existing reporting entities and 1 July 2026 for newly captured professions. The regime is technology neutral, which cuts both ways: a firm may use AI in transaction monitoring, but it must document those tools in its program, satisfy itself that the algorithms do not create detection blind spots, and keep qualified human oversight over suspicious-matter reporting.

What a board should be able to produce on request

• An inventory of every deployed and planned AI use case, with policies covering consumer fairness, bias and disclosure. This closes the gap REP 798 named.
• Evidence that the efficient, honest and fair obligation, misleading-conduct prohibitions and directors' duties have been applied to AI outputs, for example a fairness review of a credit or claims model.
• Board AI risk reporting with defined escalation triggers, and a record of directors' AI literacy training.
• CPS 230 registrations for AI vendors and foundation-model dependencies, with tested contingency plans for supplier concentration.
• Security testing results for prompt injection, data leakage and insecure integrations, and an identity regime that covers non-human and agentic actors.
• A privacy policy with the automated decision-making section published before 10 December 2026, and the internal register behind it.
• AML/CTF program documentation that names each AI monitoring tool and the human oversight around it.

Independent certification is one way to carry that evidence to a regulator, a counterparty or a tender. Responsible AI Australia's Govern tier is assessed against exactly this kind of framework, and the result is verifiable live by anyone, including the AI agents procurement teams now use.

Questions people ask