Fri, 20 Nov 2026

Readying for new risk class with agentic AI boom

By 2027, agentic AI in Asia Pacific will shift decisively from pilot to production. Enterprises in the region are deploying AI agents to triage alerts, orchestrate remediation, and accelerate software development and customer engagement alike.

But governance is not keeping pace, and the new risk core is uncomfortable in its simplicity: AI agents are becoming highly privileged "non-human identities", yet most enterprises cannot answer the most basic questions — how many agents are running, who owns them, and what data they can reach.

The stakes rise on both sides. Fortinet's FortiGuard Labs warns in its Cyberthreat Predictions for 2026 that adversaries are industrialising their operations, using automation and AI to compress the attack life cycle and force defenders to contain in minutes rather than days.

Gartner has placed agentic AI at the top of its cybersecurity trends for 2026, cautioning that unmanaged agent proliferation is creating entirely new attack surfaces.

A new identity class, with new complexities

Scott Flower

Scott Flower, senior director and head of security intelligence for APAC at Fortinet, argues that the starting point is to treat an AI agent the way an organisation treats an individual joining a company. "We need a process within the business to validate and authenticate that identity, just as we do for a human," Flower says.

An agent needs an established identity, a defined set of systems it may touch, and explicit denial of access to everything else.

What makes agents different from earlier machine identities, Flower notes, is autonomy. "Agents can access other systems, examine patterns of behaviour, analyse and aggregate those patterns to form conclusions, and potentially take actions without human intervention."

That capacity to reason, chain tasks and act elevates an agent from a service account to something closer to a digital colleague — one that still demands least privilege and compartmentalisation.

Gartner frames the same challenge as identity and access management adapting to AI agents, with new demands in identity registration and governance, credential automation and policy-driven authorisation for machine actors.

Its guidance to CISOs is blunt: "Cybersecurity leaders must identify both sanctioned and unsanctioned AI agents, enforce robust controls for each and develop incident response playbooks to address potential risks" (Alex Michaels, Director Analyst, Gartner, February 2026).

The governance playbook

Flower's prescription begins with traceability. "The first thing is to establish a unique, traceable identity for every AI agent," he says.

Around that identity sit tightly scoped, least-privilege permissions matched to the agent's purpose, time-limited access so privileges do not persist indefinitely, and continuous monitoring that preserves an audit trail for forensic and legal purposes.

The FortiGuard Labs report reaches a similar conclusion, recommending strict least-privilege and time-bound access controls for both human and non-human identities, with identity behaviour monitored across EDR (endpoint detection and response), NDR (network detection and response), SIEM (security information and event management), SOAR (security orchestration, automation and response) and CNAPP (cloud-native application protection platform) platforms.

Related:  AI and PhaaS turn inbox attacks into identity breaches

Crucially, Flower insists on human accountability. "Every AI agent needs an actual human owner who is responsible for it," he says. "It is not enough to establish identities without assigning someone accountability for them."

The dual objective: the agent must deliver a useful business outcome while its access remains constrained within defined boundaries. It is, he argues, an application of the oldest principles of identity and access management—knowing what the identity is and what it should have access to.

You cannot govern what you cannot see

Visibility is the recurring weakness. Flower points to the supply chain as one of the biggest gaps: agents may legitimately need to call on resources the enterprise does not own, so organisations must continuously monitor agent requests, tool calls, resource consumption and data transfers — both to detect unusual activity and to generate the audit trail that makes an incident reconstructable.

When something goes wrong, rehearsed procedures must contain an affected agent and revoke its credentials, "just as we would if a trusted human insider started accessing areas they should not," Flower says. Organisations, he adds, must exercise those procedures so staff know what to do before the moment arrives.

The preparation gap is acute in Asia Pacific. IDC's FutureScape research finds that only 7% of regional enterprises are highly prepared in governance, risk and compliance (GRC) skills to support tightening government controls on data and AI systems.

IDC contends that security and trust are becoming engineered, governed and continuously validated capabilities — reflected in regional security spending forecast to grow from US$39.5 billion in 2026 to US$52.4 billion by 2029. (source: IDC FutureScape: Worldwide Security and Trust 2026 Predictions – Asia/Pacific)

Shadow AI is a human problem

Unsanctioned agents compound the visibility problem. Flower is sceptical of blanket bans: businesses are pursuing AI for economic benefit, whether reduced cost or faster time to market, and prohibition forfeits that value.

"This is as much a human problem as it is an AI problem," he says. The answer is central policy across every business function, backed by training — "It is not enough to give people the tools; they need to understand how to use them properly" — and, where shadow AI has already taken root, deliberate discovery followed by central ownership, visibility and accountability.

Gartner's survey data suggests the scale of that discovery task: more than 57% of employees use personal GenAI accounts for work, and 33% admit entering sensitive information into unapproved tools.

Its recommendation mirrors Flower's — shift from generic awareness training to adaptive, AI-specific behavioural programmes.

As Alex Michaels, director analyst at Gartner, puts it: "To realise the full potential of AI in security operations, cybersecurity leaders must prioritise people as much as technology."

When agents spawn agents

The hardest questions arrive one level down the autonomy stack: what happens when an agent creates sub-agents of its own, and one goes rogue? Flower's answer returns to accountability and data stewardship.

Related:  Security in the spotlight as GITEX AI ASIA opens in Singapore

Data fed into frontier models must be properly secured and compartmentalised, and someone must be answerable for how it is used. "You cannot detach the human from the release of information," he says. "Someone makes that decision, and someone needs to be accountable for it."

Chains of delegation do not dilute responsibility; they demand it be assigned explicitly at every level.

Protecting the data journey

For data protection in hybrid environments, Flower argues the perimeter must expand beyond the model itself to "the full journey of information: how agents retrieve it, process it and share it with connected applications" — requiring "a more holistic view of the enterprise, the model itself and everything that is connected to it and that it can use."

He advocates classifying information using simple schemes such as the Traffic Light Protocol — green shareable openly, amber only within trusted networks, red never onward — then applying least privilege to the repositories agents draw on.

Mapping matters too. Flower emphasises a visual understanding of which agents touch which environments and where information lands. "Data localisation and sovereignty are increasingly important considerations," he says, and mapping what is accessed and transferred is what makes auditing possible.

This aligns with IDC's observation that Asia Pacific enterprises are reassessing cloud strategies, adopting retrieval-augmented generation and exploring private compute to satisfy data residency and sovereign AI requirements.

Security across the lifecycle

Flower argues security must begin before any agent exists, with the first question being whether the capability is needed at all: "We need a strong business use case, but also a security use case," he says, with a cost-benefit analysis conducted before deployment.

Identity, ownership, testing and integration follow in development, supported by scenario-based risk modelling. Because the knowledge to assess an AI use case rarely sits in one team, he recommends a RACI model — many may be responsible for reviewing a development, but exactly one owner is accountable.

The regional reckoning

The window is narrow. FortiGuard Labs describes a landscape in which "velocity now defines risk"—attackers operate as industrial systems, and the average annual cost of cybercrime is expected to exceed US$23 trillion by 2027.

IDC adds that adversaries are already crafting synthetic identities that blend real and fabricated data, eroding trust across financial services, e-commerce and government platforms, even as 39% of regional enterprises plan to deploy AI in security operations within the next year.

For CISOs heading into 2027, the mandate is consistent across every source: invent the identity discipline now, before the agents outnumber the humans they serve. Establish the identity, assign the owner, scope the access, log everything — and rehearse for the day an agent, or one of its children, goes rogue.

Those that treat agentic AI as an identity problem, not a technology novelty, will capture its value without inheriting a risk class they never saw coming.

Related Stories

MORE STORIES