‘You cannot secure what you don't know exists,’ says TrendAI’s Sharda Tickoo on AI agents

Indian companies giving artificial intelligence (AI) agents access to business systems must also be ready to stop them when they overstep, Sharda Tickoo, country manager for India and SAARC at enterprise cybersecurity firm TrendAI, told ET AI.

Businesses need to know which agents are operating, limit what they can access and establish who is accountable for their actions before expanding deployment, she said.

“The most important learning for India right now is to build strong governance side by side with AI adoption, before an incident forces us to catch up,” Tickoo said.


The stakes rise as AI moves beyond answering questions to accessing databases, interacting with applications and carrying out tasks with limited human intervention. In more complex arrangements, one agent can delegate work to several others, each with access to different systems and data. Delegation also makes it harder for companies to predict the steps an agent might take to complete a task.

Find the agents, then set the limits

For Indian companies, Tickoo said the starting point is identifying AI tools and agents already operating across the organisation, including those introduced without the security team’s knowledge.

A Cloud Security Alliance survey released in April 2026 found that 82% of respondents had discovered previously unknown AI agents in the preceding year, while 65% reported an agent-related security incident.

Tickoo compared the problem with early cloud adoption, when employees and business teams set up cloud instances without informing the chief information security officer (CISO). Security teams sometimes discovered them only after they had created additional entry points for attackers.

“You cannot secure what you don't know exists,” she said.

Discovery must be followed by an assessment of each agent’s permissions, the data it can access and the actions it can take. An agent performing routine administrative work poses different risks from one allowed to access sensitive financial information or initiate payments.

Tickoo said companies need guardrails that define permitted behaviour, monitoring that reveals when agents stray beyond those boundaries and mechanisms to contain them quickly. Guardrails alone may not be enough; organisations also need visibility into agent activity and mechanisms to respond when something goes wrong, she added.

Why pilots struggle to become business tools

These controls also matter when companies move from testing AI to using it in daily operations, where productivity gains must hold up against questions about reliability, security and accountability.

Tickoo said a TrendAI survey found that around one-third of AI projects did not progress beyond pilots. Demonstrating an application’s potential, she said, does not settle the questions businesses must answer before putting it into use.

A Gartner report released in January 2026 estimates that at least 50% of generative AI projects are abandoned after proof of concept because of issues such as poor data quality, inadequate risk controls, rising costs, and unclear business value.

Before deployment, Tickoo said, companies must establish who takes responsibility for errors, when humans should intervene and how agents will remain within approved limits.

These questions need to be addressed before companies deploy agents more widely, she said. A chief AI officer may guide adoption, while the CISO establishes security controls and monitoring and business leaders remain accountable for consequential decisions.

“Security should be built in. It should not be bolted on to any AI use case before we make it live, or before the organisation decides to scale it,” Tickoo said.

No model gets a free pass

Those requirements should apply regardless of the model provider, Tickoo said, particularly as Indian enterprises choose between models based on capability, suitability and cost.

A provider’s lack of reported incidents does not establish that its models are safer, she said. Incident disclosures also depend on how extensively developers test models and what they make public.

“I wouldn't infer safety just because no incidents are reported,” she said, calling for comparable testing under comparable conditions.

Companies therefore need to test what data each model receives, where that information goes, what actions the agent can perform and how quickly the provider will report an incident, she said.