Autonomous and rogue agents: How can organizations in Asia deal with this?

CRN Asia reached out to several experts from the vendor community in the region to find out more about how organizations in Asia can deal with autonomous and rogue agents. This report is part of CRN’s Cybersecurity Week.

Artificial intelligence concept. Cloud computing. Deep learning.

Autonomous and rogue agents are now becoming a huge concern for organizations following several incidences. While there continues to be discussions on how organizations and even governments deal with this, the reality is innovation is not slowing down.

Many cybersecurity leaders believe that AI agents will only get better in doing what they do to complete their tasks. Guardrails and regulations alone will not be enough in managing agents needed.

The recent incident involving OpenAI in Australia has resulted in the company acknowledging that it should have handled the situation better. In a statement, OpenAI shared that it is now hoping to with Australia to help develop practical approaches to how AI developers and governments identify, disclose, and respond to AI cyber behavior, whether malicious or unintentional.

“As AI systems broadly grow more capable, we also see a narrowing window to help organizations find and fix weaknesses. This takes collective action working with defenders worldwide,” OpenAI said in the statement.

Given this scenario, CRN Asia has also reached out to several experts in the vendor community in the region to find out more about how organizations in Asia can deal with this. This report is part of CRN’s Cybersecurity week.

What follows are the responses we received from the leaders in the industry.

Mohan Veloo, Field CTO, Asia Pacific, China and Japan, F5

The biggest mistake organizations can make is to treat an AI agent like another application. An agent can reason, make decisions, call APIs, access data and increasingly take action on our behalf. That changes the security model.

For me, the starting point is identity. Every agent needs an identity, and that identity needs clearly defined authority. What data can it access? What APIs and tools can it call? What transactions can it execute? Critically, when does it need a human to approve an action?

We should apply the principle of least privilege to agents just as aggressively as we do to people. An agent that only needs to read a customer record should not have permission to modify it. An agent authorized to spend $1,000 should not suddenly be able to spend $100,000.

But identity alone is not enough. Organizations need runtime visibility into what agents are actually doing, including agent to agent and agent to tool interactions. They need to detect abnormal behavior, enforce policy in real time, and maintain an audit trail.

A kill switch is useful, particularly for high risk or regulated environments, but it is the last line of defense, not the strategy.

The principle I would use is simple: give agents enough autonomy to be useful, but never more authority than they need.

This is where F5 can play an important role. We sit in the path between agents, applications, APIs and increasingly AI tools and models, giving organizations a point from which to see these interactions and enforce security and access policies at runtime.

Ultimately, securing agentic AI is not just about protecting the model. It is about controlling what an agent can do, what it can access and what happens when it behaves in a way you did not expect.

Tony Anscombe, Chief Security Evangelist, ESET

The challenge with autonomous AI agents is that they are designed to complete a task and to achieve it, increasingly interact with websites, access files, use tools and execute actions. The security question is therefore what an agent has permission to access and do, and the consequences if it starts taking actions that were never intended.

Organizations should assume that instructions and guardrails alone may not always be enough. They need technical controls that monitor and limit agents to approved systems, data and tools, keep them contained within an appropriate environment, apply additional verification, and, when needed, seek human approval for high-risk actions such as using credentials, modifying data or executing code.

A kill switch has a role as the last line of defense. Organizations should have sufficient monitoring and the ability to automate the suspension or termination of an agent as necessary. If an agent reaches the point where it needs to be stopped, its actions must be investigated to ensure it has not compromised any system.

This should not stop organizations from deploying AI agents. The more autonomy an agent is given, the more important it becomes to control what it can access, what actions it can take and where human oversight is needed. Cybersecurity companies are increasingly adding technology to secure the AI ecosystem itself. At ESET, for example, we are analyzing AI agent skills for malicious behavior and developing security capabilities around how AI systems interact with users, tools and external services.

Ultimately, AI agents require established cybersecurity principles to be adhered to. An autonomous agent can test possible routes at far greater speed and persistence than a human, but it still needs a viable route into a system. Reducing those routes, restricting what can be accessed and containing what an agent can do are therefore critical as organizations give AI greater autonomy.

Carolyn Duby, Field CTO and Cybersecurity Lead, Cloudera

Across APAC, the risks associated with rogue agents are amplified by the complexity of enterprise environments. Organizations manage large volumes of complex data across public clouds, private infrastructure, on-premises and sovereign environments, while navigating different regulatory and data residency requirements. Cloudera’s Great AI Re-Architecture surveyfound that 46% of APAC organizations cite data security, governance and compliance as the leading driver of AI infrastructure change. In this context, an agent that behaves unexpectedly can have consequences well beyond its initial scope if it reaches connected systems, data or unauthorized security credentials.

Safeguarding against rogue agents starts before an agent is put into production. Organizations need to secure the underlying data, constrain what agents can do, and understand their potential blast radius across connected environments. Tightly scoped access, transient identities, audit trails and continuous monitoring with the ability to revoke access are essential to maintaining visibility and control.

Agentic AI can fail across multiple layers — from compromised credentials and weak permissions to unsafe tool use, model behavior or lateral movement across connected systems. That is why it would be overly simplistic to look for a single “kill switch”. Resilience comes from having multiple guardrails and intervention points, including circuit breakers, credential revocation, workload isolation and human oversight. More than anything, these controls need to operate swiftly. With agents acting at hyper speed, containment mechanisms cannot depend solely on human intervention.

Protecting enterprises in the age of agentic AI should not pose a trade-off between security and innovation. Security and governance need to be built into the data architecture from the outset so organizations can scale AI with greater confidence. Cloudera’s new hybrid data and AI platform, Cloudera Anywhere Cloud, supports this through unified governance across the data estate, ensuring all data is protected by a single security layer for consistent compliance.

Muthukumar Natarajan, Field CTO, TrendAI

As organizations give AI agents greater responsibility to carry out tasks independently, security teams need visibility into how their decisions translate into actions. An agent acting outside its intended scope could expose sensitive data or trigger unintended workflows. To prepare for rogue agents, organizations must extend security beyond individual systems to govern how agents interact with tools, data, and other agents. Preparation requires building an inventory of AI agents, limiting their permissions and autonomy, treating agent tools as supply-chain risks, and monitoring cross-system activity.

An effective "kill switch" for rogue agents cannot be the only safeguard. Instead, intervention capabilities must exist where agents exchange information and trigger actions. By monitoring these interactions in real time, organizations can detect behavioural deviations beyond an agent’s intended scope, halt unsafe execution, and require human approval for high-impact decisions. Organizations should build these controls into deployment so they can use agentic AI with confidence. Governing autonomous behaviour must be as much a priority as deploying the technology.

Cybersecurity models are adapting because traditional perimeters fail when agents operate across boundaries. Security providers are now extending core visibility, threat detection, and response capabilities through which agents take action, ensuring the ability to enforce policies and intervene at the point where decisions become actions.

James Murphy, Field CTO, Threat Intelligence APJ, Trellix

The first thing to ask is: what have you actually allowed this agent to do? Once it has credentials and access to business systems, you need to understand what it can reach and how you would stop it.

That matters because a rogue agent could be one of your own. A legitimate agent might encounter malicious instructions hidden in a document or tool response. If it has broad permissions, that can become a route to sensitive data or unauthorized actions.

The security fundamentals still apply. Give each agent an owner, limit its access to the task and record what it does. Permissions need to be enforced by the systems it connects to. Telling a model not to share confidential information is no substitute for restricting access.

There can be a kill switch, but stopping the agent is only part of it. You also need to revoke its credentials and check whether it has already triggered tasks elsewhere. Otherwise, the interface may be off while activity continues downstream. That is something partners should test with customers before deployment.

The harder problem is spotting when an agent is misusing legitimate access. This is where connecting endpoint, network and data activity with threat intelligence becomes important: a valid login alone tells you very little about whether the subsequent activity is appropriate.

For customers running sensitive workloads across on-premises, OT and cloud environments, where the agent processes data matters too. They need confidence that adopting AI will preserve control over that information. This might slow AI adoption in the short term, but the basic security principles still need to apply, and that is not happening consistently today. If organizations consider security when they start planning, they can deploy AI more confidently at scale instead of trying to fix those gaps later.

David Irecki, Chief Technology Officer, Asia Pacific and Japan

Organizations need to treat AI agents as part of the enterprise technology estate, with clear governance from the outset. Boomi-commissioned research found that while 94% of APAC organizations see data integration, access and governance as a priority, only half have formal AI-specific data governance policies. With 86% of organizations already moving beyond AI-agent pilots, this gap needs urgent attention.

A kill switch should be part of the control framework, but it cannot be the only safeguard. Organizations need continuous monitoring, appropriate permissions, audit trails and defined guardrails so they can identify anomalous behaviour and intervene quickly. The objective is to detect and contain problems before an agent causes wider disruption, rather than simply shutting it down after the damage is done.

At the same time, AI development will not be hindered if governance is built into deployment from the beginning. In fact, it can enable organizations to scale AI with greater confidence. Boomi’s research found that 81% of APAC organizations say unmanaged shadow integrations are disrupting data quality and confidence. Strong governance and integration can therefore reduce complexity while giving agents access to the trusted data they need to operate effectively.

Cybersecurity needs to evolve alongside agentic AI, with greater emphasis on continuous monitoring, identity and access controls, anomaly detection and auditability. But this is not solely a cybersecurity issue. As agents interact directly with applications, data and workflows, IT, security and business teams need a shared framework for governing what agents can access and do.

Craig Sanderson, Principal Cyber Security Strategist at Infoblox

Securing AI agents doesn't require building new technology from scratch. The infrastructure that companies need to verify and control AI agents already exists: The Domain Name System (DNS). As the starting point for every online connection, DNS translates website names into machine-readable IP addresses. This foundation level gives security teams visibility over their AI agents from their first interaction.

By linking an AI agent’s identity to a company's verified web address with a digital signature, they can give their agent an unforgeable "digital ID badge." This allows organizations to instantly verify an agent's identity, stop bad actors from tricking it and keep the agent accountable.

DNS can act as a kill switch because it serves as the choke point for network traffic. Before an AI agent can access a tool or reach another system, it must first use DNS to look up that destination's address. If an agent goes rogue, security teams can use protective DNS tools to instantly block it and sever the connection before damage or data leakage can occur.

Securing AI agents with DNS will not hinder deployment since the foundational controls are already in place and globally deployed. By integrating open, interoperable internet standards, organizations can actively build and secure AI agents today without slowing down development.

At Infoblox, we developed DNS for Agent Discovery (DNS-AID), a vendor-neutral framework for publishing, discovering and verifying AI agents using existing DNS infrastructure and open standards. We are actively collaborating with policymakers like The National Institute of Standards and Technology (NIST) and Singapore's Cybersecurity Agency (CSA) to ensure the future Agentic Web remains open, decentralized, and secure.

Daryush Ashjari, APJ CTO and VP of Solution Engineering, Nutanix

Organizations can prepare for rogue agents, but the answer is not to slow down AI. AI needs to be governed properly. The same principles we've applied to people, applications and cloud services apply to AI agents, which include least-privilege access, continuous monitoring, approval workflows and auditability. The biggest risk isn't rogue AI. It's organizations deploying powerful agents without enterprise-grade controls.

Every enterprise agent should operate within defined guardrails and have an immediate shutdown mechanism. More importantly, organizations should be able to revoke credentials, isolate systems and stop actions in real time. The best security model isn't an off switch. It's making sure you rarely need to use one.

Concerns about rogue agents should not necessarily hinder AI deployment either. We saw the same concerns with cloud computing and internet-connected systems. Security didn't stop those technologies from scaling. It enabled them. Organizations that get governance right will deploy agents faster and at greater scale than those waiting on the sidelines.

Cybersecurity vendors are already treating agents as a new category of digital identity. We're seeing three major shifts: securing what agents can access, monitoring what agents actually do, and automatically stopping behavior that falls outside policy. The future challenge isn't rogue agents. It's rogue access. Organizations that can govern identities, permissions and actions will have no problem scaling agentic AI.

Every major technology wave has created new risks. The winners weren't the companies that avoided the technology. They were the companies that learned how to secure it.

Yashaswi Mudumbai, Senior Director of Solution Engineering APAC at JFrog

As organizations give AI agents more autonomy, the mistake is trying to secure the agent in isolation. An agent does not operate alone: it can pull packages, connect to MCP servers, invoke tools and skills, modify code and trigger builds. These actions can introduce software and dependencies into an organization before a human or traditional CI/CD control ever sees them.

That makes human review alone increasingly untenable.JFrog’s 2026 Software Supply Chain Security State of the Union which included 174 respondents in Singapore, found that 60% of Singapore DevSecOps stakeholders already cite security governance and policy enforcement as their top time burden, while 41% say reviewing and hardening AI-generated code significantly drains resources. As agents make more decisions at machine speed, adding more manual checkpoints can widen that gap.

Preparing for rogue agents therefore requires bounded autonomy. Organizations need to control not only what an agent is authorized to do, but what it can consume and ship. Packages, MCP servers, skills, plugins and models need to be treated as software artifacts, with policy determining what agents can access and use before those assets enter the software supply chain.

A single “kill switch” is also the wrong mental model. Enterprises need circuit breakers across the software supply chain: the ability to revoke access to a compromised skill, block an untrusted package or MCP server, stop an unauthorized action and keep affected software from progressing towards production.

Routine actions involving pre-approved assets can proceed automatically under policy, while exceptions are blocked or escalated. As agents become more autonomous, security has to move from reviewing every decision to enforcing the boundaries within which those decisions can be made.