AI Agent Governance Under Scrutiny by Australian Regulators
Regulatory Concerns and Findings
The Australian Prudential Regulation Authority (APRA) has highlighted significant gaps in AI agent governance within financial firms. A targeted review of large regulated entities in late 2025 revealed that while AI is widely deployed, the maturity of risk management and operational resilience varies considerably.
Key Governance Shortcomings
APRA identified several critical issues:
- Insufficient board oversight: Boards often rely on vendor presentations without adequate scrutiny of risks such as unpredictable model behaviour and the impact of AI failures on core operations.
- Lack of dedicated risk frameworks: Many institutions treat AI risk similarly to other technologies, ignoring model-specific biases and behaviour.
- Missing controls: Gaps were found in model-behaviour monitoring, change management, decommissioning procedures, and the maintenance of comprehensive AI tool inventories with clear ownership.
- Human-in-the-loop deficiencies: High-risk decisions frequently lack required human involvement, raising compliance concerns.
Cybersecurity Implications
AI adoption introduces new attack vectors, including prompt injection and insecure integrations. Identity and access management (IAM) practices have not fully adapted to non-human actors, leading to potential vulnerabilities in privileged access, configuration, and patching processes.
Industry Practices and Risks
Entities are trialling AI across software engineering, claims triage, loan processing, fraud detection, and customer interaction. However, reliance on a single AI provider is common, and few firms have formulated exit or substitution strategies.
Regulatory Recommendations
APRA advises boards to develop a deeper understanding of AI to align strategy with risk appetite, implement robust monitoring, and establish clear procedures for error handling. Specific recommendations include:
- Develop AI risk inventories with named-person ownership.
- Implement human-in-the-loop controls for high-risk decisions.
- Strengthen IAM to accommodate AI agents, including privileged access management and secure configuration.
- Adopt security testing for AI-generated code and enforce change-release controls.
Standard-Setting Initiatives
The FIDO Alliance has formed an Agentic Authentication Technical Working Group to create specifications for agent-initiated commerce, addressing the mismatch between traditional authentication models and delegated software actions.
Vendors such as Google (Agent Payments Protocol) and Mastercard (Verifiable Intent framework) are contributing solutions, while the Centre for Internet Security provides AI security companion guides mapping CIS Controls v8.1 to large language models, AI agents, and Model Context Protocol environments.
Conclusion
APRA’s review underscores the urgent need for comprehensive AI governance frameworks within the Australian financial sector. By addressing identified gaps—particularly in risk monitoring, IAM, and human oversight—institutions can better manage the evolving threats and operational challenges posed by AI integration.
