APRA Criticizes Australian Banks' AI Governance

APRA's Alert on AI Agent Governance

The Australian Prudential Regulation Authority (APRA) has recently warned Australian financial institutions about deficiencies in the governance and assurance of AI agents. This alert comes as the use of artificial intelligence becomes widespread in the internal and customer-facing operations of banks and superannuation trustees.

In 2025, APRA conducted a targeted review of large regulated entities to assess AI adoption and associated risks. While AI is present in all the entities examined, APRA noted variability in risk management and operational resilience. Boards show a marked interest in AI, particularly in terms of productivity and customer experience, but many still struggle to implement effective risk management related to AI.

Challenges in AI Risk Management

APRA expressed concerns regarding boards' excessive reliance on vendor presentations, without sufficiently considering risks such as the unpredictable behavior of AI models and the potential impact of failures on critical operations.

The agency emphasized the importance for boards to develop a deep understanding of AI in order to define coherent strategies and oversight. These strategies must align with the institution's risk appetite and include procedures for managing errors.

Entities are testing or introducing AI in various areas, including software engineering, claims processing, loan processing, fraud detection, and customer interaction. However, treating AI risk like that of other technologies does not account for the specific behaviors and biases of AI models.

APRA identified gaps in monitoring model behavior, change management, and decommissioning. It stated that it is necessary to establish inventories of AI tools and designate responsible parties for each instance of AI. APRA also clarified that AI could be present in upstream dependencies that entities were not aware of.

APRA highlighted the need for human involvement in high-risk decisions to ensure that critical actions do not rely solely on automated systems.

Cybersecurity and Identity Management

Cybersecurity is another area of concern. APRA noted that the adoption of AI is changing the threat landscape, adding attack vectors such as command injection and insecure integrations. Identity and access management practices are not always suited for AI agents.

APRA recommends strict controls on autonomous workflows, including privileged access management and the security of AI-generated code. The reliance on a single vendor for many AI instances is also concerning, with few entities having an exit or substitution plan.

The volume of AI-assisted software development puts pressure on change and release controls, necessitating particular attention to maintain the security and efficiency of systems.

FIDO Alliance Initiatives

Meanwhile, the FIDO Alliance is working on new standards for identity and permission controls, forming an Agentic Authentication Technical Working Group. This group is developing specifications for agent-initiated commerce.

FIDO emphasizes that current authentication models are designed for human interaction, requiring adaptations for actions delegated by software. Solutions like Google’s Agent Payments Protocol and Mastercard’s Verifiable Intent framework are being examined to meet these needs.

The Centre for Internet Security has published guidelines on AI security, addressing prompt and sensitive data issues, and focusing on secure access through software tools and non-human identities.