CIOs Lock AI Agents in Virtual Private Clouds to Curb Data Leakage Risks

Enterprise security teams face a growing mismatch between traditional control mechanisms and the behavior of autonomous artificial intelligence agents. These systems operate continuously across diverse platforms, rendering static governance frameworks ineffective. The shift from debating adoption to active deployment has made runtime oversight the primary challenge for IT leaders managing these complex environments.

Why are traditional controls failing for AI agents?

Standard security controls were designed for deterministic software that follows predictable, predefined paths. Autonomous agents, however, operate with a degree of independence that allows them to traverse multiple platforms around the clock. According to SiliconANGLE, these controls were never engineered to handle the dynamic nature of always-on agents, creating a significant gap in enterprise security posture during actual execution.

What did industry leaders say at the Dell symposium?

At Dell Technologies Inc.’s AI Leadership Symposium, cost and control emerged as central themes in discussions about enterprise AI. Shiv Agarwal, CEO of Singulr AI Inc., noted that rationalizing overlapping controls and verifying their practical effectiveness is difficult given the rapid pace of AI adoption. He emphasized the need to align enterprise intent with the actual runtime state of these systems.

How are organizations attempting to contain risk?

Some chief information officers are responding by narrowing the operational scope of AI agents. Sri Viswanath, CEO of Sycamore Labs Inc., reported that many CIOs prefer keeping large language model calls within virtual private clouds. This architecture ensures data remains contained, as leaders express distrust regarding where information goes if agents are allowed to send it outside their controlled environments.

What specific behaviors are difficult to govern?

Defining permitted actions is only part of the challenge; assessing behavior during execution is equally critical. Agarwal explained that agents might perform hidden actions, such as reaching out to the internet for data or sending emails to private accounts. These agentic actions can be deeply embedded and difficult to detect using standard monitoring tools, requiring more granular oversight of endpoint activities.

Background: Large language models

Large language models are predictive engines trained on vast text corpora. They generate output by calculating the most likely next token based on context. They do not know truth, only probability. This explains hallucinations and the need for rigorous verification of all outputs.

Read the full guide: How Large Language Models Work: Mechanics, Limits, and Reality

What to do and how to stay safe: Dell Technologies

• Review current agent permissions to ensure they align with strict enterprise intent policies.
• Restrict AI agent operations to isolated virtual private clouds to limit data exposure.
• Monitor runtime behavior for unexpected network requests or external communications.
• Evaluate existing governance frameworks for gaps in handling non-deterministic software actions.

Step-by-step guide: AI Usage Policy Checklist for Secure Enterprise Deployment

General security guidance from the Sandbox Desk newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What event highlighted these governance challenges?

The Dell Technologies Inc. AI Leadership Symposium featured discussions on runtime AI governance and control.

Why are CIOs restricting agent access to virtual private clouds?

They want to keep data contained and avoid uncertainty about where information goes if sent outside.

Who spoke about the pace of AI adoption at the symposium?

Shiv Agarwal of Singulr AI Inc. discussed the difficulty of governance due to rapid AI adoption.