Why Singapore is most vulnerable to the rising wave of AI security risks – Security #AI

Singapore is seeing a sharp and structural rise in cybercrimes. As organisations deploy AI agents, the question is no longer about an external breach. The challenge is to govern what an enterprise’s own autonomous systems are permitted to do.
 Can an organisation’s policies anticipate its AI’s next move and trust it?
To explore why Singapore organisations are experiencing a high rate of AI-security challenges and to get insights on how to counter these threats, iTNews Asia, speaks with Andrew Kay, Senior Director of Systems Engineering at Illumio, APJ.
According to Kay, an autonomous agent with legitimate credentials and live connectivity, can move through an enterprise faster than the organisation’s policies can anticipate. This turns access into a potential attack surface.
What’s happening in Singapore on AI security
Greater exposure, better detection, weaker controls, are amongst the many reasons why Singapore organisations report AI security incidents at a higher rate than most global peers.
Singapore leads APAC on AI governance, but enforcement has not yet caught up, research by Illumio found. While the region’s organisations are well informed about AI risk, exposure, however, remains a larger and more serious issue.
Singapore enterprises have extended AI significant access. Among those willing to accept more risk, more than half of Singapore organisations are enabling AI to act autonomously in production and four in 10 assign it privileged identities, both well above the global average in Illumio’s report. A similar number have found an AI agent or tool connected to systems without security approval, the highest rate of any market we surveyed. “That is not a detection effect. It is AI reaching systems that no one authorised,” the research cited.
Singapore organisations are more cautious than most about taking on additional risk. The gap lies in enforcement.
Illumio’s report found nine in 10 Singapore organisations experienced at least one AI-related security issue in the past year, the highest of the 10 markets in its research. Only about 4 in 10 organisations can see and audit every action an AI agent takes, compared with 49 percent globally.
On the contrary, Australian organisations have embedded AI more deeply in the security function than any other market, with four out of five using it in cybersecurity operations. Japan sits at the more cautious end of the spectrum. Fewer than half of Japanese organisations have delegated decisions to AI, and they are three times more likely than the global average to prioritise risk reduction over new AI benefits.
Hong Kong’s critical infrastructure cybersecurity law is nonetheless requiring operators in banking, energy, transport and healthcare to map their critical systems and dependencies. That same discipline is essential for AI.
“The common thread across every market is clear. The authority organisations give AI is outpacing their ability to see and constrain what it does. Singapore is simply experiencing that gap earlier than most,” said Kay.
AI has moved from adviser to operator and that shift has happened faster than most security programmes were designed to accommodate, Kay further added.
Lingering concerns on the use of agentic AI
AI adoption is accelerating beyond security. The Economic Survey of Singapore shows AI adoption is concentrated in information and communications, electronics, professional services, finance and insurance.
Financial services warrants particular attention. Banks and insurers are among the most advanced adopters, and they are also typically the most rigorous on traceability.
Sumsub’s APAC AI Governance Benchmark, produced with the Singapore FinTech Association, ranks financial services first on AI governance, with a region-leading 68 percent of firms maintaining audit trails of AI decisions. IT and software services follow closely, although the report warns that their rapid deployment could outpace governance.

The concern is that the fastest-moving sectors often operate the most interconnected environments, spanning hybrid cloud, legacy core systems and third-party integrations. Each of those connections is a potential path for an agent. Rapid adoption creates real value. However, without enforced boundaries minor error escalates into a serious breach.   
–  Andrew Kay, Senior Director of Systems Engineering at Illumio, APJ

Who is accountable when things go haywire
Accountability rests with the organisation. A task can be delegated to an agent, but responsibility for its actions cannot.
IMDA’s Model AI Governance Framework for Agentic AI is clear on this point: while agents may act autonomously, human responsibility continues to apply. It states that the organisations deploying agents — and the humans overseeing them — remain accountable for agents’ decisions and actions.
The increased capability and autonomy of agents also create challenges for effective human accountability, such as greater automation bias, or the tendency to over-trust an automated system that has performed reliably in the past. It is therefore crucial to understand the risks agentic AI could pose and ensure that organisations implement the necessary governance measures to harness agentic AI responsibly, including maintaining meaningful human control and oversight over agentic AI, reads the framework, according to Model AI Governance Framework For Agentic AI.  
“Every agent in production should have a named owner who can answer three questions immediately: what it is permitted to access, what it has actually accessed, and how to shut it down. Without those answers, accountability exists in name only. Policies fail at the point of use because most are written around intent rather than reach,” Kay said.
The distance between principle and practice is also significant.
“The solution is to move policy from documentation into the environment itself. When boundaries are enforced within the application and across the underlying network, policy is no longer guidance an agent may or may not follow. It becomes a constraint the agent cannot work around,” Kay added.
He said that when an agent operates within enforced boundaries, a missed alert or a skills gap does not automatically become a breach. The impact stays contained.
Kay recommends several practical measures organisations can look at to address the human gaps.
•       A shared view of the environment. Security, IT and AI teams need a single, consistent picture of what connects to what. Threat correlation becomes far more effective when every team works from the same map.
•       Security from day one. No agent should reach production without a defined scope and security sign-off. Four in 10 Singapore organisations have already seen agents bypass that step.
•       Measure reach, not alert volume. The meaningful question is how far a compromised agent could move, and how quickly the organisation could stop it.
How enterprises can become resilient
AI has changed the economics of attack. Exploits that once required skilled teams and weeks of effort can now be developed in hours.
The top priority should be containing the blast radius, said Kay.
Autonomous agents can probe for weaknesses, exploit legitimate access paths and move laterally across connected systems at machine speed, he added.
He explains that this makes detection essential, but it cannot be the only line of defence as no security team can identify and stop every attack in real time. Some intrusions, unfortunately, will succeed.
“What matters is what happens next. An organisation’s resilience depends on whether it can limit an attacker’s ability to move beyond the initial point of compromise, isolate affected systems quickly and prevent a localised incident from becoming an enterprise-wide breach,” Kay said.
According to him, resilient organisations prepare for that outcome before an attack occurs, rather than trying to improvise containment during one. They map how their systems, workloads, identities and AI agents connect, then restrict communications so that a single compromise cannot spread unchecked across the business. They also ensure that affected systems can be isolated in seconds, rather than days.
This principle applies equally to external attackers and to AI already operating inside the organisation, Kay added.
In both cases, he said the most effective response is to enforce boundaries they cannot cross and microsegmentation provides the practical mechanism for doing so. 
“The guidance is straightforward. Organisations should look beyond keeping attackers out and ask how far an attacker could get once inside the system. The goal is to make that distance as small as possible, measure it consistently, and to keep reducing it,” explained Kay.
 


Click Here For The Original Source.

——————————————————–

............