Australian Firms Hand AI Agents the Keys, but Many Cannot Audit Every Move, Illumio Finds

Australian companies are giving artificial intelligence agents real power inside their systems, and a new research report suggests oversight has not kept pace. The study, titled Licensed to Act: How AI Autonomy Is Outrunning the Limits Placed on It, was conducted by Sapio Research for Illumio and surveyed 1,100 senior IT security decision makers across 10 markets, including Australia. Its message is plain: AI is no longer just answering questions. In many workplaces it is now making decisions, changing settings, and reaching into sensitive systems, sometimes without a human checking first.

The headline numbers are striking. According to the report, 80 per cent of Australian organisations now delegate decisions and actions to AI that required human approval a year ago. Yet only 55 per cent can see and audit every action taken by AI agents. An agent acting with legitimate access looks, on the surface, like normal activity, which makes an unexpected action much harder to spot than a break in from outside.

Australia leads on AI in security work

Australia stands out for how far AI has spread inside security teams themselves. The report says 80 per cent of Australian businesses use AI within cybersecurity operations, the highest share of any market surveyed and well above the global average of 67 per cent.

That adoption has moved quickly from assistance to autonomy. In Australia, many organisations now allow AI to take autonomous action without human review, including granting or revoking access permissions in 33 per cent of cases, running security investigations in 28 per cent, and deploying code changes unsupervised in 26 per cent.

Incidents are already showing up

The risks are not only theoretical. Almost three quarters of Australian businesses, 73 per cent, experienced at least one AI related security issue or warning sign in the past 12 months, against 76 per cent globally. In Australia, 35 per cent of IT security decision makers report that AI accessed data it should not have, and 33 per cent say AI made a recommendation that would have created a security risk if followed. Another 30 per cent flagged agents or tools connecting to systems without security approval, and 28 per cent flagged an unauthorised action.

Globally, the most common incident was a tool recommending security actions that would have created new risk exposure if followed, reported by 35 per cent. Behind that sits a cluster of boundary failures: agents connecting to systems without approval at 26 per cent, accessing data they should not have at 24 per cent, and triggering unauthorised actions at 24 per cent.

Confidence is running ahead of visibility

One striking finding is the gap between what leaders believe they can see and what they can actually demonstrate. The report says 84 per cent of IT and security leaders say they have clear visibility into how AI systems access data and systems. Yet when asked what their team can actually observe, fewer than half, 49 per cent, say they can see and audit every action an agent takes. Another 32 per cent see most actions but not all, 12 per cent see only outcomes rather than individual actions, and 6 per cent have limited or no visibility at all. That is a 35 percentage point gap between claimed visibility and demonstrated visibility.

Confidence in guardrails runs high at the same time. Nearly all respondents, 92 per cent, are confident their internal guardrails would stop a model acting outside its intended role, and 89 per cent say the same about their providers' guardrails. Yet 48 per cent say they are not prepared for the risks introduced by frontier models.

Organisations are knowingly accepting more risk

Rather than pulling back, many organisations are trading more risk for the benefits AI brings. Globally, 69 per cent of decision makers agree their organisation has accepted higher levels of risk to realise the benefits of AI, and 56 per cent say their organisation accepts more AI related cyber risk than it did 12 months ago. Over a third now issue privileged identities to AI, giving it elevated credentials normally reserved for administrators. The report also finds that 62 per cent agree they deployed AI tools before establishing appropriate safeguards, a pattern it compares with the early years of cloud adoption.

The practical response: limit reach, keep logs

The Australian Signals Directorate has weighed in with advice that lines up with the report. Christopher Horlyck, Head of Cyber Security Resilience at the Directorate, said the findings reflected a real and growing risk, and warned that agents with excessive access or insufficient controls may take unintended or unauthorised actions. Its guidance advises restricting access to what is necessary, maintaining monitoring and audit logs, and applying human oversight to high impact actions.

Illumio's response is built on Zero Trust principles. Rather than trying to predict every action an agent might take, organisations should explicitly define the identity, access, connections, and actions each agent needs, and limit what it can reach beyond those boundaries. Andrew Kay, Senior Director of Systems Engineering, APJ at Illumio, said Australian organisations are putting AI in the security driver's seat, but many cannot see or audit every action it takes. Australia already leans this way: 37 per cent of Australian organisations say frontier AI has increased the importance of network segmentation, compared with 27 per cent globally. The next step is turning that instinct into controls that hold even when an agent does not cooperate, starting with an inventory, least privilege access, permitted actions, and logs a human can actually review.

Frequently Asked Questions

What is an AI agent in this context?

An AI agent is a software system that can take actions on its own to pursue a goal, such as changing access permissions or updating code, rather than only answering a question when asked. The Illumio report focuses on agents that operate inside business systems with legitimate credentials.

Why is auditing AI agents harder than auditing people?

An agent with legitimate access can look like normal activity, and its actions may be logged only as outcomes rather than step by step decisions. The report found that only 49 per cent of teams globally can see and audit every action an agent takes.

What should an Australian business do first?

Start with an inventory of every AI agent and tool in use, then give each agent its own identity with the smallest permissions it needs, define exactly which actions it is allowed to take, and keep logs that a human can review.