Australia’s warning about AI agents arrives at an awkward moment: decades of underinvestment have left core public-sector and utility systems reliant on brittle, interdependent platforms. These environments weren’t designed for intelligent, persistent automation probing every exposed interface. As AI agents grow more autonomous—looping through discovery, exploit selection, and privilege escalation—the economics tilt further against legacy operators, especially where maintenance windows are politically unpalatable and downtime is reputationally costly.
The threat is not only zero-days. Agents supercharge the mundane: password spraying tuned by behavioral signals, MFA fatigue blended with context-aware social engineering, and precise chaining of misconfigurations that human operators often miss. In mixed IT/OT estates, outdated protocols, flat networks, and long patch cycles widen the blast radius. Unsupported vendor stacks and bespoke integrations frustrate basic hygiene, while data-rich SaaS connectors and unmanaged secrets provide low-friction entry and durable persistence for automated adversaries.
At the policy edge, Australia is debating AI guardrails and sovereign capabilities, yet resilience hinges on clearing security debt before layering in new AI services. Procurement must reward decommissioning, not only acquisition, and require observable architectures—asset inventories, SBOMs, identity baselines—that let defenders measure risk reduction. Sovereign AI should mean dependable compute, trusted data pipelines, and incident response muscle memory, not just model hosting. Without this reframe, the country risks importing sophistication while exporting stability.
For practitioners, the near-term playbook is ruthless prioritisation. Map crown-jewel dependencies, kill exposed legacy services (RDP/SMB/old VPNs), and segment OT from corporate IT with strict egress controls. Elevate identity—phishing-resistant MFA, hardware-backed keys for admins, and continuous access evaluation. Invest in detection that understands automation loops: high-fidelity telemetry, eBPF-based sensors, and model-assisted anomaly triage. Validate with red/purple teaming and tabletop exercises that assume autonomous, tool-using adversaries. Budget line-items should show measurable reductions in time-to-mitigate and attack surface, not just new platform logos.