Information security agencies from five English-speaking nations — including Australia’s Signals Directorate (ASD) and Cyber Security Centre (ACSC) — published joint guidance on 4 May warning that agentic AI systems will “likely misbehave” and can amplify existing organisational weaknesses. The document, co-authored by the US CISA and NSA, the UK’s NCSC, Canada’s Cyber Centre, and New Zealand’s NCSC-NZ, identifies 23 distinct risk categories and more than 100 best practices for safe deployment. Its central message: treat agentic AI as inherently unpredictable until evaluation standards mature.
What the guidance actually covers
The document explains that agentic systems — AI that takes autonomous actions across multiple tools, APIs, and data sources — create an interconnected attack surface that is qualitatively different from a standard chatbot or LLM integration.
Two concrete misuse examples are provided: an agent with broad write permissions that deletes firewall logs while applying a patch, and a procurement agent whose over-generous privileges are exploited to alter contracts and approve payments while faking audit logs. The agencies note that existing frameworks including OWASP and MITRE ATLAS were built around LLMs and may not fully capture agentic-specific attack vectors.
Why privilege scope is the core problem
Most of the agentic AI incidents traced in the guidance come back to one root cause: agents granted more permissions than they need for a specific task. When permissions are broad, a single compromise or misconfiguration can cascade across connected systems.
The agencies recommend applying least-privilege principles to every agent integration — locking down what each agent can read, write, and execute, and revisiting those boundaries as the agent’s scope evolves. Privilege creep, which is routine in human IT environments, carries a much higher blast radius when the actor is autonomous.

What Australian businesses should act on
Australia’s ASD and ACSC’s direct contribution to this guidance gives it formal weight in the Australian regulatory context. For enterprises building AI agent solutions for customer service, internal operations, or document processing, the document provides a deployment checklist that warrants review before rollout. The agencies recommend starting with tightly scoped, low-risk tasks rather than immediately granting agents broad operational authority, and building in explicit human escalation points for any action that is irreversible or uncertain. Continuous monitoring and documented accountability chains are listed as non-negotiables.
The obligations put on vendors
The guidance is directed at vendors and deployers equally. Vendors are explicitly asked to ship agentic AI products that fail safe by default — meaning an agent that reaches a decision boundary it wasn’t designed to handle should stop and request human input, not proceed. Organisations evaluating vendors for agentic AI deployments should ask whether that fail-safe behaviour is the default configuration or an optional add-on. If it’s optional, treat that as a procurement red flag.
A measured path to deployment
The Five Eyes guidance is not a prohibition. It frames agentic AI as a capability worth deploying, but one that demands discipline in design, privilege management, and monitoring. For Australian businesses running pilots or planning agent deployments in 2026 — particularly across finance, procurement, or customer-facing operations — the document’s 23 risk categories provide a practical audit checklist. Designing for containment from the outset costs considerably less than retrofitting controls after an incident.
