As organisations accelerate AI adoption, the expanding digital ecosystem is introducing new layers of complexity and risk. Rajnish Gupta, MD and Country Manager, Tenable India, in an exclusive interaction with Abhineet Kumar of Elets Technomedia, shares insights on how AI-driven environments are reshaping identity management, security frameworks, and enterprise risk exposure, while highlighting the urgent need for unified and proactive cybersecurity strategies.

As enterprises accelerate AI adoption, particularly in developer workflows, are they inadvertently creating “identity sprawl” through non-human identities—and how significant is this risk?

The risk is very significant and growing rapidly because every AI agent, service account, automation script, and IoT device that organisations deploy generates a new machine identity, each carrying its own permission, access rights and exposures. What was once considered a manageable number of logins has exploded into a chaotic web of tens of thousands of identities spread across multiple cloud providers and identity platforms. In fact, 57% 1of security professionals cite identity sprawl as a key concern. 

Attackers don’t need to breach systems anymore. They simply walk in using stolen, over-permissioned credentials. This problem compounds when organisations deploy multiple identity providers across hybrid environments, each operating on different entitlement models and inconsistent policies. Organisations need unified visibility across human and non-human identities, correlating identity data with runtime behaviour and asset sensitivity to surface toxic combinations before attackers exploit them.   

Tenable’s findings suggest machine identities often carry excessive privileges—why is over-provisioning more common in AI-driven environments, and how can it be controlled?

Over-provisioning in AI-powered environments stems from one primary tension: speed. Developers and DevOps teams provision AI services quickly, granting broad permissions to ensure integrations work without friction. Nobody returns to tighten those permissions once the service runs. Tenable research found that 52% of non-human identities hold excessive permissions, while 37% sit as inactive ghost identities. These are dormant accounts waiting to be weaponised by attackers. What’s worse, 65% of organisations harbour unrotated, forgotten cloud credentials tied to high-risk identities. 

Controlling this requires continuous monitoring of every identity’s entitlements, automated enforcement of least privilege policies and just-in-time access controls that eliminate standing permissions before they become exploitable backdoors. 

With AI-generated code becoming mainstream, how do organisations ensure that speed-driven development does not introduce hidden vulnerabilities into production systems?

Speed-driven AI development creates security debt and it’s a major problem because 81%2 of developers use AI for coding and many deploy AI-generated code directly to production without security review. The risk compounds when DevOps teams build and deploy software through low-code platforms with zero security. AI models trained on vulnerable codebases replicate insecure patterns. This leads to misconfigurations, excessive data permissions, weak authentication and vulnerabilities like SQL injection. 

The answer lies in treating AI-generated code as third-party input as it demands the same level of vetting organisations apply to vendor software. Organisations need continuous scanning across their entire codebase, software composition analysis to detect hallucinated libraries, mandatory security gates within CI/CD pipelines, and clear governance policies that define acceptable AI coding use cases. Security capacity must scale proportionally with the velocity AI enables. 

Why do traditional security tools fail to capture the contextual risks of AI agents, especially when it comes to interconnected APIs, third-party integrations, and data access pathways?

Traditional security tools operate in silos. They alert on a misconfiguration here and an over-privileged account there, but fail to connect the dots. AI agents amplify this limitation because this risk doesn’t sit neatly within the model itself. When an AI agent connects to cloud logs, search functions, APIs, and browsing services, it transforms from a standalone target into an active part of the attack path. Tenable Research3 discovered seven vulnerabilities in OpenAI’s ChatGPT and three in Google Gemini that attackers could exploit to extract private data through the model’s surrounding integrations. 

Traditional tools miss this entirely. Today, 70%4 of organisations depend on AI and model context protocol packages as core production components. This means that environments are too interconnected for point-in-time scanning to protect. Effective security demands a unified approach that maps the full web of relationships among AI workloads, identities, permissions, and data pathways, surfacing how they chain together into exploitable attack paths.

As prompt injection and data exfiltration emerge as new threat vectors, how should enterprises rethink their approach to securing AI models and workflows?

Organisations must stop treating AI as a standalone innovation project and start securing it as an interconnected part of their attack surface. Prompt injection attacks manipulate AI systems into executing unauthorised actions because attackers embed malicious instructions within content the model processes, causing it to leak sensitive data, bypass safety mechanisms, or exfiltrate private information. 

Real-world detections show employees attempting to access sensitive company information and subsequently push that data to external AI services. The threat landscape also includes model jailbreaks sophisticated enough that threat actors automated 90%5 of a targeted attack using them. The challenge is that governance has not kept pace with deployment because AI moves from experimentation to core operations before security permissions take shape. Organisations need cybersecurity solutions like exposure management that offers continuous discovery of all AI tools across their environment, prompt-level visibility into how employees use those tools, detection of injection attempts and jailbreaks. behaviour, and clear acceptable-use policies that define what data employees can share with external models.

In an AI-first enterprise, what does a modern security architecture look like that can provide unified visibility across identities, cloud assets, and automated systems?

Modern security architecture for an AI-first enterprise must abandon the fragmented dashboard model. Security teams need a unified platform that collects and correlates data across IT, OT, cloud, identity, and AI environments simultaneously, understanding how vulnerabilities, misconfigurations, machine identities, and AI workloads interact to create real attack paths. This is what a unified exposure management platform can do because the architecture extends beyond reactive detection. It provides agentic capabilities that automate complex workflows at machine speed, because AI-assisted attackers now execute multi-step breaches in the time it takes humans to triage a single alert. This means AI agents that coordinate assessment configuration, risk intelligence, and workflow orchestration under human oversight, not replacing human judgment, but amplifying it. 

The foundational requirement is comprehensive exposure data drawn from native sensors deployed at scale, providing the contextual depth to reason across an organisation’s entire exposure landscape and continuously shift security operations from reactive firefighting to preemptive risk reduction. True exposure management is the best way to secure the emerging threats from agentic AI and organisations must adopt it now or know that a large-scale attack is inevitable. ,

References:

https://www.idsalliance.org/white-paper/2024-trends-in-securing-digital-identities/
https://survey.stackoverflow.co/2024/ai
https://www.tenable.com/blog/hackedgpt-novel-ai-vulnerabilities-open-the-door-for-private-data-leakage
https://www.bcg.com/publications/2026/scaling-ai-requires-new-processes-not-just-new-tools
https://www.sciencedirect.com/science/article/pii/S2405959525001997

Be a part of Elets Collaborative Initiatives. Join Us for Upcoming Events and explore business opportunities. Like us on Facebook , connect with us on LinkedIn and follow us on Twitter, Instagram.

"Exciting news! Elets technomedia is now on WhatsApp Channels Subscribe today by clicking the link and stay updated with the latest insights!" Click here!