Singapore organisations are reporting AI-related security incidents at a higher rate than any other market surveyed. The findings suggest that, as enterprises delegate more decisions and actions to AI agents, many have yet to establish sufficient visibility over what those systems can access and do.
The Illumio report, Licensed to Act: How AI Autonomy Is Outrunning the Limits Placed, found that 91% of Singapore organisations had experienced at least one AI-related security issue or warning sign in the past 12 months, compared with a global average of 76%.
The research also highlights a widening gap between the authority organisations grant AI and their ability to audit it.
Among respondents from Singapore, 74% now delegate decisions or actions that would have required human approval a year earlier, but only 43% say they can see and audit every action taken by an AI agent.
Unapproved access emerges
Unapproved connectivity stood out as a regional concern with 41% of respondents from Singapore rating an AI agent or tool had connected to systems without security approval, highest, well above the 26% global average.
A further 35% reported that AI had produced or triggered an unauthorised action, while 30% said AI had accessed data it should not have reached.
Illumio reflects on this as a new form of insider risk where an organisation’s own AI may already possess legitimate credentials and network access, allowing it to reach sensitive systems and data if controls are poorly defined.
The trend is particularly relevant as AI moves deeper into security operations. Among Singaporean organisations, three-quarters use AI within cybersecurity operations, while 30% are comfortable allowing AI to conduct security investigations without human review.
The country has urged organisations to assess and bound risks early by limiting agents’ autonomy and their access to tools and data. It also calls for meaningful human accountability, technical controls throughout the AI lifecycle, and transparency for end users.
Despite this, the Illumio research indicates that practical enforcement remains inconsistent. Only 39% of Singapore respondents said they were very confident they could contain a compromised or misused AI system before it reached critical systems or data, compared with 57% globally.
Although 72% agreed that limiting AI access matters more than predicting every threat, only 25% identified explicitly defined permitted actions as one of their two main controls.
“Singapore has done more than most countries to set out how agentic AI should be governed, and organisations here are clearly careful about the risk they take on,” said Andrew Kay, senior director of systems engineering, APJ, at Illumio.
He warned that caution alone is not going to keep AI in its lane.
Andrew Kay
"When four in 10 organisations have found AI plugged into systems nobody signed off on, the problem isn’t intent. It’s reach.” Andrew Kay
He argued that businesses should define precisely what each AI agent can touch and enforce those limits across applications and underlying infrastructure, rather than attempting to anticipate every action an agent may take.
For Singapore enterprises, the research reinforces the direction set by IMDA: constrain an agent’s reach through Zero Trust principles and enforceable access controls, rather than relying solely on prediction, policy documents or human oversight after the fact.
