Australian firms let AI act with limited oversight

Illumio has published research showing that 80% of Australian organisations now let artificial intelligence make decisions and take actions that required human approval a year ago. The study also found that only 55% can see and audit every action taken by AI agents.

The figures come from a global survey of 1,100 senior IT security decision-makers, including respondents in Australia, and point to a widening gap between the authority businesses give AI and the controls they retain.

Australia stood out for its use of AI in cybersecurity operations. The research found that 80% of Australian businesses use AI in cybersecurity, the highest share among the markets surveyed and well above the global average of 67%.

That use is also extending to autonomous tasks. One-third of Australian organisations said they allow AI to grant or revoke access permissions without human review, while 28% let it run security investigations and 26% permit unsupervised code changes.

The report suggests many companies are running into problems as AI systems gain broader access to applications, data and infrastructure. Nearly three-quarters of Australian businesses, or 73%, said they had experienced at least one AI-related security issue or warning sign in the past 12 months.

Among the issues reported, 35% said AI had accessed data it should not have. Another 33% said AI made a recommendation that would have created a security risk if followed, while 30% said AI tools or agents connected to systems without security approval and 28% said AI produced or triggered an unauthorised action.

A senior Australian cyber official said the findings underline the need for tighter controls.

"As AI is used more extensively in cybersecurity operations, Australian organisations must ensure this does not come at the expense of security and control. AI agents with excessive access or insufficient controls may take unintended or unauthorised actions," said Christopher Horlyck, Head of Cyber Security Resilience, Australian Signals Directorate.

"The Australian Signals Directorate's recent guidance, Agentic AI harnesses and the Information Security Manual, provide practical advice on managing AI risk, including restricting access to what is necessary, maintaining monitoring and audit logs, and applying appropriate human oversight to high-impact actions," Horlyck said.

Risk appetite

The research also suggests some organisations are becoming more willing to accept cyber risk in exchange for the perceived benefits of AI. More than half of respondents globally, 53%, said they were willing to accept more cyber risk than they were 12 months ago.

Within that group, more than half allow autonomous AI actions in non-production environments. A further 45% allow those actions in production environments, where systems are live and errors can have more immediate operational effects.

The survey indicates that many security leaders believe AI needs different controls from conventional software. Eighty-four per cent said AI agents require fundamentally different security controls, while 77% said limiting what AI can access matters more than predicting every threat.

Even so, the controls needed to define those limits do not yet appear to rank highly. Only 37% said frontier AI had made network segmentation or microsegmentation more important, and 26% named explicitly defined permitted actions as one of their two main controls.

Illumio said this mismatch leaves organisations exposed because AI systems can act with legitimate access rather than breaking in from outside. That can make problematic behaviour harder to detect in environments where unusual activity may resemble routine automated work.

"Australian organisations are putting AI in the security driver's seat, but many can't see or audit every action it takes. That's a dangerous gap. Unlike external hackers, AI agents can operate with legitimate access to sensitive systems and data, making their actions harder to distinguish from normal activity," said Andrew Kay, Senior Director of Systems Engineering, APJ, at Illumio.

"As Australia leads the world in AI adoption within cybersecurity, visibility and control must keep pace with autonomy. AI agents are relentless in pursuit of their goal, by design. Security teams cannot predict every action an agent might take, so they must limit what it can reach and be able to contain the damage when it behaves unexpectedly," Kay said.