Most identity programmes in Australian business were built around people: staff, contractors and the occasional service account. The software now running inside those businesses does not fit that model. AI agents hold credentials, reach into live systems and act without a person watching each step. Palo Alto Networks found that 91% of Australian organisations experienced an identity breach in the past 12 months, according to ARN.
The argument is narrower than it first sounds. It is not that agents are inherently unsafe. It is that the controls most organisations already run, built for employees who join and leave, were never designed for software that acquires its own access and quietly keeps it.
What the report actually found
The figure comes from the 2026 Identity Security Landscape report, which Palo Alto Networks used to argue that identity security has moved from covering people to covering machines as well. The company positioned its Idira platform as a single place to handle privileged access management, machine identity security and AI identity governance.
That framing serves the vendor, and it is worth reading with that in mind. The underlying number is harder to wave away. An identity breach rate of 91% across Australian organisations in a single year suggests the problem predates AI agents rather than being created by them. What agents change is the rate at which new credentials appear.

Why agent credentials slip through
AI agents accumulate access quickly. They connect to code repositories, cloud infrastructure, CRMs and payment platforms, and every connection needs a credential. Those credentials are usually issued mid project, by whoever happened to be building it, and they rarely carry an expiry date.
When a person leaves an Australian business, something happens. Payroll knows, IT knows, and the accounts get closed. When a pilot project ends, the agent that ran it often keeps its API keys, its database access and its permissions, because nobody owns the decision to remove them. The credential outlives the use case, and it does so silently.
Why Australian businesses should care
The regulatory position is clearer than most teams assume. APRA CPS 234 requires regulated entities to control access to information assets, and CPS 230 extends that expectation to operational risk and service providers. The Essential Eight treats restricting privileged access as a core mitigation. The Security of Critical Infrastructure Act applies to a widening set of operators.
None of those frameworks distinguish between a human login and a machine one. An agent with write access to a finance system is an access path, and an auditor will treat it as one. The Department of Industry has published its own guidance on risks and controls for multi agent systems, which is a fair signal of where local expectations are heading.
It also overlaps with a question Australian firms already ask. An agent holding credentials is another path for data to leave the country, which puts it beside the same concerns covered in our note on data residency. The access question and the sovereignty question turn out to be the same question wearing different clothes.
Where to start looking
The first pass does not need a platform purchase. It needs a list. Which agents are running, who created them, what credentials they hold, and what those credentials can reach. Most organisations that run this exercise find at least one agent nobody remembers approving, still authenticated against a production system.
After that it is ordinary access hygiene applied to something new. Scope each agent to the narrowest permission set that still works, give its credentials an expiry, and log what the agent actually does rather than what it was meant to do. That last point matters, because an agent’s behaviour drifts as its prompts, tools and models change underneath it.
None of this is exotic. It is the part of AI strategy work that rarely makes the business case and reliably turns up in the audit.
A quieter kind of exposure
This is not a dramatic risk. There is no single incident to point at, and the failure mode is slow: a spreading set of forgotten keys, each one reasonable at the moment it was issued. It sits close to the governance gap we covered in June, where the controls existed on paper and nobody had mapped them to what was actually running.
For Australian businesses adding agents faster than they are adding oversight, the useful question is not whether the tools are secure. It is whether anyone could produce that list today if an auditor asked for it.
