- AI identity governance is the enterprise security gap Malaysia hasn’t budgeted for.
- Non-human identities could outnumber human users 500 to one, and most organisations have no idea what’s already running inside their networks.
Malaysian enterprises have spent the last few years deploying AI into their operations. What many haven’t done is ask what credentials those AI systems are running on.
Every AI agent, every automated workflow, every machine-to-machine integration deployed inside an enterprise needs an identity to operate: a credential that grants it access to systems, data, and processes. And right now, most Malaysian enterprises are provisioning access the same way they did a decade ago: quickly, broadly, and without much thought about what happens when the project ends.
“The most common architectural gap we observe is the tendency to treat autonomous AI agents as if they are standard, static application integrations,” says Jay Reddy, Head of Growth at ManageEngine, in an email interview with Tech Wire Asia. “Developers routinely provision these autonomous agents with standing administrative permissions or over-privileged service accounts simply to avoid day-to-day integration bottlenecks.”
The result is an attack surface that is expanding faster than anyone is tracking.
AI identity governance: the gap nobody budgeted for
ManageEngine’s Identity Security Outlook 2026 puts a number to the problem: non-human identities could soon outnumber human users by as much as 500 to one, driven heavily by the rapid deployment of enterprise AI agents. That’s a significant jump from Gartner’s 45-to-one baseline ratio for typical hybrid-cloud environments, and the gap between the two figures is largely explained by one thing: orphaned credentials that nobody is managing.
When a human employee leaves an organisation, there’s an offboarding process: HR coordinates, access is revoked, accounts are closed. Machine identities have no equivalent. An AI agent provisioned for a temporary migration or a localised model test gets its credentials, completes the task, and then quietly persists in the background–fully privileged, unmonitored, and in most cases, completely forgotten.
“You cannot secure or manage what you do not know exists,” Reddy says, “and at a 500-to-one scale, this unmapped entitlement sprawl creates severe structural vulnerability.”
A significant portion of this volume is driven by shadow IT. Business units bypass central IT governance to deploy low-code automation, macro scripts, or specialised AI tools, each spinning up new clusters of service accounts and API keys without any decommissioning plan. When security teams are running fragmented, legacy point solutions, they lack the visibility to even discover these connections, let alone govern them.
Gartner forecasts that 40% of enterprise software applications will feature task-specific autonomous AI agents by the end of this year. When four in ten applications are launching autonomous operations inside a network, continuing to provision them with unmanaged standing access isn’t just risky; it’s architecturally unsustainable.
What an attack actually looks like
Human-centric security has improved considerably. Multi-factor authentication, phishing awareness training, and tighter endpoint controls have made targeting human workers more visible and more costly for attackers. So they’ve moved to the back channels.
“In contrast, machine-targeted credential abuse is entirely technical and completely silent,” Reddy explains. Attackers scan cloud environments to hijack active tokens, scrape orphaned API keys left exposed in development code, or manipulate AI model parameters to force an agent to execute unauthorised actions, all without triggering the user behaviour alerts that traditional security tools are built to catch.
Once inside, the attacker exploits the blanket trust that enterprise environments extend to internal machine-to-machine communications. The credential is already recognised as a trusted internal service. Lateral movement across data silos happens quietly, often for extended periods, before anything flags.
MyCERT’s Cyber Incident Quarterly Summary Report for the fourth quarter of 2025 noted a rise in unauthorised access incidents resulting in enterprise data breaches in Malaysia, reflecting what Reddy describes as a deliberate strategic shift by threat actors away from front-door attacks and toward automated back channels.
Governing AI agents without grinding everything to a halt
The standard objection to tighter identity governance is developer velocity. If every AI agent needs manual approval and sign-off, deployment slows to a crawl, and the efficiency gains that drove AI adoption in the first place evaporate.
Reddy answers that the friction itself is a symptom of outdated thinking. “Governance historically meant forcing manual approvals that slowed processes to a crawl,” he says. “We must shift toward fully automated, risk-aware identity frameworks.”
The practical mechanics involve two things. First, every non-human credential should carry a Time-to-Live (TTL) metric: automated systems should never hold permanent, standing privileges. Access should default to expiration, requiring continuous automated context justification to remain active. Second, every line of autonomous code should be cryptographically signed and mapped to an accountable human or system owner at runtime.
The governance question sits directly alongside something ManageEngine’s own leadership raised when Tech Wire Asia spoke to CEO Rajesh Ganesan in Jakarta last month: that the enterprise AI conversation has been almost entirely about capability and deployment speed, and not nearly enough about what happens operationally once AI systems are inside the network and making decisions. Identity is part of that reckoning.
Zero Trust needs to be rebuilt for machines
Zero Trust security frameworks–never trust, always verify–are widely cited in enterprise cybersecurity conversations in Malaysia. The problem is that most Zero Trust implementations are designed around human signals: location, biometrics, working hours, device fingerprints. An AI agent has none of those attributes.
“What has to be rebuilt from scratch is the engineering underneath it,” Reddy says. Non-human Zero Trust requires continuous ambient identity validation: authenticating the code itself, confirming an agent hasn’t been tampered with or injected with a malicious prompt, and verifying its cryptographic integrity and historical behavioural intent rather than checking whether it logged in from the right office.
Critically, it also requires eliminating standing privileges. A Just-in-Time (JIT) access model means an AI agent gets the exact minimum access needed for a single micro-task, and the moment that task is completed, even if it takes a fraction of a second, those privileges are purged. ManageEngine’s own report finds that over 80% of security leaders are already consolidating fragmented identity tools into unified platforms to close the visibility gap between human and non-human access management.
Some security leaders push back on this approach, arguing that deploying predictive AI models to monitor other AI agents simply trades one black box for another, introducing non-deterministic risk into the SOC rather than resolving it. Reddy doesn’t dismiss the concern. “The concern raised by risk officers is completely legitimate in principle,” he says. “By nature, predictive AI models operate on likelihood rather than certainty.” But at a 500-to-one identity ratio, he argues, human analysts are mathematically incapable of monitoring the volume and velocity of machine-to-machine transactions in real time.
The alternative to imperfect AI-driven oversight isn’t perfect human oversight; it’s no oversight at all. Gartner lists AI-amplified security operations as a defining trend for modern threat mitigation, and ManageEngine’s position is that the technology acts as a filter, synthesising millions of machine events down to the few that require a human decision, keeping the CISO in the loop without requiring them to be everywhere at once.
But technology solutions can only go so far without policy backing them up.
The regulation gap
Malaysia has made real progress on AI governance at the policy level; the National AI Roadmap, NACSA’s cybersecurity frameworks, and evolving PDPA enforcement all point in the right direction. What’s missing is any specific mandate governing how enterprises manage AI agent identities or machine access controls.
Without one, the calculus for most organisations is simple: speed wins. “Relying on voluntary best practices is fundamentally insufficient when enterprises are competing in a market that aggressively rewards commercial speed,” Reddy says. “Without a formal compliance mandate, identity governance will inevitably be pushed aside in favour of rapid deployment.”
His ask of policymakers is specific: move beyond data privacy frameworks and begin designing risk-aware compliance standards that explicitly mandate lifecycle management and cryptographic accountability for non-human identities. In practice, that means incentivising platform consolidation, getting enterprises off fragmented, brittle security toolsets and onto unified, fully auditable identity strategies.
Malaysia is building AI infrastructure at pace. The governance conversation has mostly been about models, data, and ethics. The question of what credentials those AI systems are running on, and who is watching when they go dormant, hasn’t had nearly enough airtime.
The post Malaysia’s AI rollout is outpacing identity security, ManageEngine warns appeared first on TechWire Asia.