What happened
On 10 September 2026, ASD's Australian Cyber Security Centre released guidance titled 'Agentic AI Harnesses' (with accompanying PDF 'The layer above the model'), treating the software harness connecting AI models to enterprise data/tools/systems as a critical governance and security surface. It sets out expected controls: least-privilege access, strong identity and access management, secure design, monitoring/audit logging, human oversight of high-impact actions, supply-chain assurance, and phased deployment.
Why it matters
This is authoritative national cybersecurity-agency guidance (from a Five Eyes member) establishing the harness/tool-execution layer — not the model itself — as the primary control point for agentic AI risk management, shaping expectations for how Australian government and regulated entities should architect and govern agentic AI deployments.
Action needed
Organizations deploying agentic AI in Australia (especially government and critical infrastructure) should benchmark their agent architectures against ASD's control list: inventory harnesses/connectors/tools, enforce least privilege, require approval gates for high-impact actions, and validate supply chain of harness components.