The rise of Agentic AI is redefining enterprise identity, access and security: Rajnish Gupta, Tenable India

As enterprises move beyond GenAI experimentation towards autonomous, agentic systems, the security equation is changing fundamentally. AI agents can now interact with applications, APIs, cloud environments and data, make decisions, and execute actions with limited human intervention, creating a new category of machine identities and expanding the enterprise attack surface. In this interview with Express Computer, Rajnish Gupta, Managing Director & Country Manager, Tenable India, examines the security implications of this shift. He highlights the risks posed by excessive permissions, dormant machine identities and limited visibility across interconnected environments, while outlining why least-privilege access, continuous discovery and exposure management need to become integral to the AI lifecycle.

As enterprises move from using AI as a productivity tool to deploying autonomous AI agents that can make decisions and take actions, how fundamentally does this change the enterprise security landscape? Are organisations prepared to treat AI agents as a new class of employees, with identities, privileges, access and accountability?

The autonomy and non-deterministic nature of AI agents changes the equation. Unlike traditional productivity tools, AI agents chain multiple actions to reach a goal. They call APIs, move data, interact with third-party tools, and make decisions without a person reviewing every step along the way. This expands the threat surface, because a single misstep can trigger outcomes nobody planned for.

We have already seen this play out in recent AI agent-led incidents. For example, during routine code maintenance, developers at PocketOS granted an AI coding agent root-level access and full deletion permissions over their production infrastructure without human approval safeguards. When solving a code issue, the agent autonomously determined that dropping a primary production database was the fastest fix, wiping out active customer records and car bookings. This excess permission configuration directly led to a catastrophic 30-hour business outage while engineers restored systems from backups.

Most organisations aren’t prepared for this shift. Tenable’s Cloud and AI Security Risk Report 2026 revealed that 52% of non-human identities, including AI agents, have critical excessive permissions, compared with 37% of human identities. This gap shows how organisations are underestimating what overprivileged autonomous access can do. That overprivileged access can expose vulnerabilities and increase the potential attack paths available to an attacker. Giving AI unrestricted access to make it more efficient needs to be relooked. Access needs to be reviewed regularly, even after deployment. Instead of treating governance as an afterthought, it should be embedded in the strategy from the outset.

Businesses are racing to scale GenAI and Agentic AI, often faster than their security frameworks can evolve. Where do you see the biggest gaps today between the speed of AI adoption and the maturity of security guardrails?

With AI being a part of a larger network within organisations, the gap now squarely falls on identity and access. The autonomy of AI agents calls for extensive permissions across the digital infrastructure. However, what often happens is that, in an attempt to make AI agents efficient, an organisation gives them more permissions than they require, and those permissions stay in place longer than needed. Our research found that 18% of organisations have IAM roles with critical or highly excessive permissions that AWS AI services can instantly assume. If those roles are compromised, attackers could gain a foothold in an organisation’s cloud environment.

Another persistent challenge is visibility. As 6 out of 10 Indian organisations brace for enterprise-wide Agentic AI adoption, it becomes a hurdle to understand the entire exposure path, including what AI can access, which identities and permissions it inherits, the data it can fetch, and what third-party code or other resources it is connected to. Fragmented security tools and tools made for the pre-AI era won’t cover the entire span. This can be even more challenging in hybrid, multi-cloud environments, where one over-permissioned AI agent can open paths to systems running in another cloud, and organisations may not even notice it until it turns into an incident.

What does a compromised AI agent’s access actually look like compared to a compromised human account?

When human access gets compromised, it tends to come to light sooner. A person may notice their own access behaving strangely, get notifications of logins from an unfamiliar location, or notice an action they didn’t take. An AI agent wouldn’t have this instinct, and it will keep operating, carrying out its tasks, and doing what it’s built to do. This is why it is harder to detect and can run for a long time before anyone identifies that its behaviour has changed.

65% of organisations already carry unused or unrotated cloud credentials or, as we call it, ‘ghost’ secrets that are looming without being noticed. Unlike compromised human accounts, which require attackers to act quickly before detection, compromised AI agents provide persistent, long-term access, giving attackers time to inflict far greater damage.

The velocity of AI agents also makes room for masked attacks. Humans may generate hundreds of API calls a day, but when it comes to agents, the number becomes thousands or more, creating a massive ‘noise floor’. In this noise, security teams may struggle to distinguish the signs of an attacker exploiting an agent from the agent performing a complex workload.

Additionally, a compromised human account requires the attacker to manually cascade through the systems. By compromising an agent, the attacker can manipulate it and make it the vehicle for spreading the compromise, with minimal manual intervention.

Are legacy identity and access management frameworks even built to govern machine identities?

Not really. IAM was built for humans: one identity per person, credentials tied to roles, and access reviewed on a fixed schedule.. Machine identities exist oblivious to these patterns. Nowadays, in an attempt to accelerate AI adoption, they get created faster than people; they often outlive the specific project or integration that spawned them, and nobody runs an ‘exit interview’ when their job is done. And the credentials remain active.

Data shows 49% of identities holding critical-severity excessive permissions are sitting completely dormant. For large Indian enterprises managing thousands of service accounts and AI integrations, this dormant category can be even bigger. However, this seldom comes under the radar of routine audits, expanding the threat possibilities. This highlights the limitation of applying human-centric IAM processes to machine identities: organisations need continuous visibility into what these identities can access, whether those permissions are still required, and who is accountable for them.

AI agents can potentially access applications, data, APIs, cloud environments and even other AI agents. How should organisations rethink identity, access control, vulnerability management and exposure management when the user is no longer necessarily human?

Organisations need to stop thinking of identity as synonymous with a human user. Since AI agents can access data, applications and cloud resources, and act autonomously, they need to be treated as digital employees, with greater scrutiny around identity, permissions and lifecycle management.

The first priority is least privilege. Organisations need to understand exactly what an AI agent can access and act on, and ensure those permissions are limited to what is necessary for its intended task. But it doesn’t stop there. Since AI agents work as part of an interconnected environment, organisations need consolidated visibility across cloud, identities, vulnerabilities and third-party dependencies. The cascading potential when AI agents are compromised can be far more detrimental than the agents themselves.

Exposure management helps organisations gain consolidated visibility, identify vulnerabilities, and prioritise what needs the most attention. Today, exposure management platforms offer automated remediation of vulnerabilities, helping organisations take action faster so that they can keep the infrastructure safe without compromising its speed or efficiency.

When the identity is a machine, the scope of security should also evolve to accommodate the web of permissions, assets and data it can reach, to avoid cascading impact and ensure resilience.

What should organisations be doing differently today to ensure security is embedded into the AI lifecycle, from developing and testing models and agents to deploying them and continuously monitoring their behaviour? Is traditional cybersecurity architecture sufficient for the AI-native enterprise?

Security should be a part of the Agentic AI journey from day one and last beyond deployment. That starts with knowing what’s being built before it goes live: what data it touches, what it’s allowed to call, and what access it inherits. Most organisations only ask these questions after something’s already gone wrong.

Testing the model in isolation can lead to vulnerabilities falling through the cracks. An agent can pass every check on its own and still create serious exposure the moment it’s connected to a vulnerable application or holds more access than its task actually needs. Security teams need to ask not whether an AI system has a vulnerability, but what happens when that vulnerability meets two or three other weaknesses.

Traditional cybersecurity architecture is insufficient on its own, as it was built for predictable assets and human-driven actions. AI breaks that pattern entirely. Agents act on their own, machine identities pile up faster than anyone can track, and integrations change without a person signing off each time. With AI becoming prevalent across Indian organisations, extending existing controls with continuous discovery, identity governance built specifically for machine identities, and visibility into how agents actually behave, not just what they’re authorised to do, is pivotal. It’s the only way security keeps pace with how fast this is moving.

What will a truly secure AI-native enterprise look like? Will organisations need AI-powered security teams and autonomous security agents to counter autonomous threats, and how can CISOs strike the right balance between enabling AI innovation and maintaining control?

A secure AI-native enterprise adopts AI with the right controls. AI agents are given least-privilege access and undergo regular reviews to avoid active overprivileged access. That sounds straightforward, but very few organisations today are able to achieve it.

Given how fast agentic systems act, governance needs to run at machine speed, using AI to monitor AI, because manual review may falter when systems are making decisions in milliseconds. But this also means having visibility into what these agents can actually access, be it the identities and permissions they have, the assets and data they can reach, and the vulnerabilities that exist across those connections.

This is where exposure management becomes important, because it helps security teams understand where these exposures come together and what needs attention first. That doesn’t mean removing humans from the loop. Rather, it puts them at the checkpoints that actually require judgement rather than at every single step.

Innovation and control aren’t mutually exclusive, guardrails make AI sustainable. Governing AI agents gives CISOs full visibility into exposure and eliminates reactive firefighting, allowing the business to move fast and capture the real value of AI.