Briefing
Reference 14 · Primer · About 6 minutes
Download the MP3 · Linda, an AI voice, reads a script written from the sources below.
Written from Home Affairs and ASD material and practitioner summaries, checked on 11 November 2026. The ISM control wording comes from secondary summaries. Check the current text before relying on it.
Before we record, every claim the hosts make is checked against the sources (this table was last updated 11 November 2026). Confirmed means a primary source or two independent reports; Reported means one party says so; Partly right, Disputed, Unverified and Not yet known mean what they say, and the hosts say so on air. This is the table as we left it.
The rules that apply to AI in Australia and for Australian organisations with European exposure, and the practical work each one calls for.
These are Agile Insights AI Strategy & Governance offerings. Which of them you need depends on where your organisation starts.
Sign up for our podcast at agile-insights.com.au.
Linda The security rules for government AI changed several times in 2026. The newest change gives agencies a deadline at the end of March 2027. By the end of this episode, you will know the two rulebooks involved and what each now asks. You will also know what a security assessor will ask an AI supplier.
Linda First, what they are. The Protective Security Policy Framework, the P S P F, belongs to the Department of Home Affairs. It is mandatory for non corporate Commonwealth entities, and it reaches suppliers through their contracts. Underneath it sits the Information Security Manual, the I S M, from the Australian Signals Directorate. The framework sets the obligations. The manual is the catalogue of technical controls, and it is updated every quarter.
Linda The problem they solve is that AI widens the attack surface while also speeding up attackers. An agency with old, unpatched systems is exposed either way. Picture an AI agent that can read email and call other tools. Suppose it has no identity of its own and keeps no log. Then nobody can say who acted, or stop it quickly.
Linda Who uses them? Chief security officers and chief information security officers own the obligations. Security assessors under the Infosec Registered Assessors Program, called I RAP, test systems against the manual. Procurement teams write the framework into contracts. And suppliers of AI products meet both when they go through an assessment.
Linda So what is in them? Start with the framework. Since its 2025 release, agencies must put AI technologies through technology system authorisation, wherever they are hosted. A 2025 direction barred DeepSeek products. The 2026 release adds training on AI use. In May 2026, Home Affairs issued an advisory on the frontier AI era. Its message was blunt. You do not need the most advanced AI to stay protected. Get the basics right first, starting with patching and hardening.
Linda Then, in September 2026, came a binding direction on AI enabled risks. As reported, agencies must take stock of their legacy technology and write a plan to manage its risk. They must also set reduction targets and patch critical flaws faster. The stocktake and plan are due by the end of March 2027. That lands beside the AI policy's own dates, so security and AI teams will compete for the same people.
Linda Now the manual. Industry summaries of the June and September 2026 updates report a new set of AI controls. Each AI agent gets its own identity, separate from people and from other agents. Agents sit in a register with an owner, a purpose and the tools they can reach. A tool call needs both the user's access and a task specific grant for the agent. Content an agent retrieves stays untrusted. Every tool call is logged centrally. And a human approves high impact actions before they run.
Linda How does this connect to the rest of the stack? The D T A's AI Reference Architecture says it will defer to existing guidance on security, without naming it. For government, that guidance is the framework and the manual. The national Cyber Security Strategy sits above both. Its second horizon, launched in June 2026, adds national AI work such as crisis planning for major AI incidents.
Linda Getting started. First, start the legacy stocktake now, because March 2027 is close. Second, give every AI agent its own identity and an entry in a register with a named owner. Third, if you supply AI products, prepare evidence for those manual controls before an assessor asks. The primary sources are on protective security dot gov dot A U and cyber dot gov dot A U. Both are linked in the show notes.
Linda Now, currency. As at November 2026, the framework is at its 2026 release, and the manual's latest update is from September. The manual changes every quarter, and the control wording here comes from industry summaries. Check the current text before you rely on any detail here.
Linda In short, government AI security starts with the basics. Then every agent gets an identity, a log and a human brake. If you'd like to know more, sign up for our podcast at agile-insights.com.au.
All of the reference series (16) · Follow AI Governance Reference in Apple Podcasts or any podcast app
Everything from Agile Insights: briefings, Making You Think and the reference series