The AI model isn’t the whole system
When organisations assess agentic AI, most attention still goes to the model. New guidance from the Australian Signals Directorate (ASD)makes an important point: the model is only part of the risk.
ASD uses the term “harness” for the software layer connecting an AI model to organisational data, tools, memory and systems.
That distinction matters. A chatbot produces content. An AI agent may search files, access customer information, call APIs, update records or initiate actions. The harness determines what it can access, what it is permitted to do and when a person must intervene.
ASD also warns that prompt injection cannot be reliably addressed within the model alone. Controls must therefore extend across the harness, connected systems and organisational governance.
Before deploying agentic AI, leaders should be asking:
What data, tools and systems can the agent access?
Are permissions limited to what the task genuinely requires?
Which actions need human approval?
Can decisions and actions be monitored and audited?
Who remains accountable if something goes wrong?
This is where AI governance moves beyond policies and principles. It becomes a practical exercise in access control, system design, monitoring, supplier assurance and clearly defined decision rights.
Hash Elevens helps organisations make that shift. We support AI inventories and risk assessments, procurement reviews, access and accountability mapping, human-oversight design, and governance controls that can be demonstrated to boards and assurance teams.
Agentic AI can create real value. But as its ability to act increases, the controls surrounding it must become stronger, not optional.
ASD guidance: https://www.cyber.gov.au/business-government/secure-design/artificial-intelligence/agentic-ai-harnesses
Comments