Australian Signals Directorate issues security guidance for Agentic AI

The Australian Signals Directorate (ASD) has issued new guidance regarding Agentic AI Harnesses, focusing on the security, governance, and operational risks associated with these systems. The guidance emphasizes that organizations must look beyond Large Language Models (LLMs) and focus on the software layer—the harness—that connects models to organizational data, tools, and systems.

An agentic AI system integrates an LLM with external tools, data sources, memory, and planning workflows. While the LLM proposes actions, the harness provides the infrastructure to execute those actions and enforce permissions. The ASD notes that significant risks, such as prompt injection, cannot be solved by the model alone and require additional controls within the harness and connected systems.

The ASD describes the harness as a critical, long-term organizational investment that allows for direct governance and control. Key components of the harness include the policy layer, context manager, tool registry, and permission systems, which together manage access, execution privileges, and activity monitoring.