Cybernion
← Insights

IRAP

ISM AI Agent Controls: Identity, Registers and Agentic AI

4 Sept 2026 · Updated 4 Sept 2026

The September 2026 ISM treats an AI agent as an identity, not a feature. It must have a unique account, appear in a register, run on least privilege, and be prevented from trusting what it reads on the internet. Where an agentic tool has been connected to a system that holds classified data, ISM-2133 to ISM-2159 now describe how it has to be governed, and an IRAP assessment will test against them.

Which ISM controls cover AI agents?

The September 2026 release added two clusters. The first, ISM-2133 to ISM-2135, gives an AI agent an identity and puts it on a register. The second, ISM-2156 to ISM-2159, constrains what an agent can do once it is running. They build on the AI application controls from earlier in 2026, ISM-2112 to ISM-2117, and the amended human approval control ISM-2113. The ISM September 2026 changes added them alongside a broader block on non-human identity, which is the wider context these agent controls belong to.

The shift is worth naming plainly. Until this release the ISM’s AI controls treated AI as an application that gets deployed. The agent controls treat AI as an actor that authenticates, holds credentials and takes actions on its own. That is a different threat model, and it is the one that matches agentic tools.

Why does an AI agent need its own identity under the ISM?

A shared identity hides who did what, which is why ISM-2133 requires each AI agent to have a unique identity, distinct from the user accounts of staff and from other agents. The reason is accountability. If an agent runs under a person’s account or a shared service account, an agent’s action cannot be told apart from a human’s in the logs, its permissions cannot be scoped separately, and it cannot be revoked without breaking something else. A unique identity per agent makes the agent a first class subject: it can be authorised, monitored and switched off on its own.

This is where teams trip. An agent wired in through an existing automation account inherits that account’s access, which is almost always broader than the agent needs. A unique identity is the precondition for every other control in this set.

What is an AI agent register under the ISM?

The register is a maintained inventory of every agent and what it can reach. ISM-2134 requires an AI agent register to be built, implemented, maintained and regularly verified, and ISM-2135 sets out what each entry holds.

Register field What it records
Unique identifier The agent’s own identity
Owner and business purpose Who is accountable for it and why it exists
Assigned identities The identities the agent operates under
Accounts and credentials Any user accounts and credentials it uses
Tools, permissions, data The tools, permissions and data repositories it can access

Read the last row against reality. Most organisations cannot currently answer, for a given agent, which tools it can call and which data it can read. The register forces that answer, and it is the artefact an IRAP assessor will ask for first.

Does the ISM require prompt injection defence for AI agents?

In effect, yes, through ISM-2158. It requires external content retrieved by an agentic AI application to be treated as untrusted throughout processing, clearly delimited from system instructions, subject to the same validation and sanitisation as other untrusted input, and prevented from modifying or overriding system instructions, security policies, access controls, tool permissions or human approval requirements. That is a description of prompt injection defence written in control language. The attack is simple: hide an instruction in a web page or a document, let the agent read it, and have the agent act on it. ISM-2158 says the retrieved content can never become the instruction.

The rest of the agentic set reduces the blast radius if something does get through. ISM-2156 restricts an agent to the minimum tools, functions and permissions for its purpose. ISM-2157 limits a tool invocation to the lesser of the invoking user’s access and the agent’s task scoped authorisation, so an agent cannot do more than either the user or its task allows. ISM-2159 requires every tool invocation, external request and output to be centrally logged in enough detail to support an incident investigation.

Do the ISM AI agent controls apply to an IRAP assessment?

They apply if an AI agent is inside the assessment boundary and the classification selects them. Every ISM control carries an applicability marking across non-classified, OFFICIAL: Sensitive, PROTECTED, SECRET and TOP SECRET, so the assessment tailors the set to the system in front of it. If there is no agentic AI in scope, these controls are marked not applicable with a justification, and the assessment moves on.

The catch is the same one that caught teams with the earlier AI controls. An agent added after the boundary was agreed is a material change, not a quiet upgrade. A support agent that can read a ticketing system and call an internal API has an identity, a set of tools and a data reach, which is exactly what ISM-2133 to ISM-2135 want documented. Before a next IRAP assessment, running the baseline through Cybernion Local shows which of the new AI and identity controls fall inside scope.

Frequently asked questions

Which ISM controls cover AI agents?

The September 2026 ISM added ISM-2133 to ISM-2135 for AI agent identity and the AI agent register, and ISM-2156 to ISM-2159 for agentic AI restrictions. Earlier releases added ISM-2112 to ISM-2117 and amended ISM-2113 for AI application behaviour and human approval.

What is an AI agent register under the ISM?

ISM-2134 and ISM-2135 require a maintained register listing, for each AI agent, its unique identifier, owner and business purpose, the identities assigned to it, any user accounts and credentials it uses, and the tools, permissions and data repositories it can access.

Does the ISM require prompt injection defence for AI agents?

In effect, yes. ISM-2158 requires external content retrieved by an agentic AI application to be treated as untrusted, kept separate from system instructions, validated and sanitised, and prevented from overriding system instructions, security policies, access controls or human approval requirements.

Do the ISM AI agent controls apply to our IRAP assessment?

They apply if an AI agent falls inside the assessment boundary and the classification selects them. If there is no agentic AI in scope, they are marked not applicable with a justification. Adding an AI agent to an assessed system is a material change.


Written by Gaurav Vikash, an ASD endorsed IRAP assessor and senior cyber security leader with 18 years of experience across Australia, the UK and Asia, including CISO and senior security leadership roles. He holds CISSP, CISA, CISM and CRISC and is an ISO 27001 Lead Auditor and ISO 42001 Lead Implementer, and speaks regularly at industry conferences.

Talk to us. We aren’t always chasing a transaction.

Sources:

  1. ISM September 2026 changes document, ASD, September 2026
  2. Information Security Manual, ASD, September 2026
  3. Cloud assessment and authorisation, ASD, 2026

Last updated: 4 September, 2026

Related serviceIRAP AssessmentSee the service →

Talk to us about your engagement

A scoped proposal within one business day.