28% of Australian firms found AI agents nobody had registered

By DataForge team17 November 20265 min read

ServiceNow says 28% of Australian firms found unregistered AI agents in production. NZ firms should build an agent register before scaling agents.

What ServiceNow found

On 12 November, ServiceNow released the Australian results of its study Deployed but ungoverned: the state of agentic security, which it ran with Phronesis Partners. The survey covered more than 3,700 senior security and technology decision makers in 20 countries, including 300 in Australia. Of the Australian respondents, 28% said they'd discovered agents working in live systems that nobody in the business had registered or even knew about. Only 51% of the Australian sample said they use agentic AI at all. Among organisations that actually run agents, the share with unknown agents is probably higher than 28%, but ServiceNow hasn't published that breakdown.

The rest of the Australian figures point to the same gap in oversight. According to the ServiceNow release, 46% of respondents have real-time monitoring of their agents and 39% have incident response that covers agents. Only 27% have a mandatory automated gate that stops an agent going live until it passes security checks. Another 63% can't act automatically when an agent behaves abnormally. When an agent incident happened, 43% of Australian organisations failed to contain it within 24 hours, compared with 36% globally. The release also says 91% had an agentic security incident in the past year. ServiceNow hasn't published how it defined an incident, so a blocked prompt-injection attempt could count the same as a real data leak.

The global release from 21 October shows the scale behind those numbers. Of organisations surveyed worldwide, 49% run between 100 and 999 agents and 35% run more than 1,000. Only 57% keep a full, accurate record of every production agent, and just 19% have live visibility of more than 90% of their agents.

ServiceNow sells agent governance tooling, including what it calls an AI kill switch, so the study also works as marketing. The dates of the survey fieldwork haven't been published either. Separate research gives similar results. Illumio's survey, reported on 4 November, found that 30% of Australian respondents had seen AI tools or agents connect to systems without security approval, and only 55% could see and audit every action their agents take.

How agents end up off the books

Building an agent used to need a project, a budget and someone from IT. Now a staff member can switch on an agent feature in a SaaS product the company already pays for, or connect a low-code workflow to a language model in an afternoon. Agents built this way usually run under the creator's own login and inherit everything that person can reach. When the creator changes roles or leaves, nobody tells the agent. Vendors are also adding agent features to existing subscriptions and turning some on by default. In those cases an agent arrives with a routine product update and nobody inside the business made a decision to deploy it.

So the register gap is mostly a procurement and process problem, and ServiceNow's figures fit with that. When 57% of Australian organisations bring in security only after an agent has already been planned and designed, as the Australian release reports, an agent that skipped the planning stage completely never meets security at all.

What the findings mean for NZ businesses

The study has no New Zealand data. The Australian release names ServiceNow's ANZ vice president, but every figure in it is Australian. NZ organisations buy software from the same vendors as their Australian counterparts and face the same pressure to show AI gains. We'd expect a New Zealand survey to look much the same, and we wouldn't assume NZ firms are in better shape until someone measures it.

New Zealand already has a reference point. On 1 May 2026, NCSC NZ co-published Careful Adoption of Agentic AI Services with the Australian Signals Directorate's ACSC, CISA and the NSA in the US, and the Canadian and UK cyber centres. According to Cybersecurity Dive's summary, the guidance advises keeping agents on low-risk, non-sensitive tasks for now, avoiding broad access to sensitive data or critical systems, and requiring human approval for high-cost actions such as deleting records or resetting systems. A business can only apply those controls to agents it knows about. An agent nobody has recorded has no assessment of its access and nobody approving its actions. If that agent touches customer records, the business is still accountable for the result.

For organisations that run IT service management, the fix fits into processes they already have. An agent behaves like a configuration item that can take actions. Recording agents alongside servers and applications, and sending each new agent or vendor-enabled agent feature through change management, gives the agent an owner and a review date. Firms that already follow ITIL 4 practices won't need a new governance body. They need to extend the scope of their existing configuration and change records.

Practical next steps

  • Run a discovery sweep. Check the admin consoles of your main SaaS platforms for agent features that are switched on, and review the OAuth app grants and API keys issued in the past 12 months. Ask each team lead what automations or assistants their people have connected to company systems.
  • Start a register, even in a spreadsheet. For each agent, record its owner and business purpose, the identity it runs under, the systems and data it can reach, whether it can write or only read, and the date it was last reviewed.
  • Give agents their own identities. An agent running under its own service identity can be cut off without locking out a staff member, and the logs will show which actions the agent took.
  • Gate production. Make an owner and a register entry conditions of go-live, and treat a vendor's default-on agent feature as a change that needs assessment before it's left running.
  • Define the actions that need sign-off. Follow the NCSC co-authored guidance and require human approval for deletions, payments, permission changes and anything else that's hard to reverse.
  • Test the off switch. Confirm that someone on call can disable each agent, and record how long it took. Add agent ownership to your staff offboarding checklist.

ServiceNow's numbers come from a vendor with something to sell, but the underlying problem is easy to check inside your own business. A discovery sweep takes days, and it'll show whether you have one unknown agent or dozens. If you're planning to use more agents in 2027, our AI agents page sets out how we approach design, permissions and oversight from the start.

Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.

Get my free assessment How we help
Get new articles by RSS