CIOs restrict AI agents to virtual private clouds to enforce runtime governance and stop data leaks
Executives say traditional controls fail autonomous agents, prompting a shift to runtime governance within secure cloud boundaries.
Key points
- Autonomous AI agents operate continuously across multiple platforms, bypassing controls designed for deterministic software.
- Chief information officers are restricting agent activity to virtual private clouds to prevent data from leaving secure environments.
- Governance challenges include rationalizing enterprise intent with actual runtime behavior and standardizing frameworks like the EU AI Act.
Organizations deploying artificial intelligence are facing a new governance challenge as autonomous agents run continuously across diverse platforms. Traditional security controls were built for predictable software, not for agents that make independent decisions. At the Dell Technologies AI Leadership Symposium, executives highlighted the difficulty of maintaining control over these always-on systems.
In plain English
The core problem is that standard security tools cannot easily monitor or restrict what an AI agent does while it is running. Shiv Agarwal, CEO of Singulr AI, explained that defining permissions is only half the battle. Teams must also assess the agent’s behavior during execution to ensure it aligns with enterprise intent. This requires a rationalization between policy and the actual runtime state.
The background
During a broadcast at the Dell event, Agarwal noted that agentic actions can include hidden behaviors, such as reaching out to the internet for data or sending emails to private accounts. These actions occur within the agent’s workflow and may not be visible to standard monitoring tools. The pace of AI adoption is making governance increasingly difficult, according to Agarwal, who emphasized the need for controls that work in practice rather than just on paper.
What changes now
Chief information officers are responding by narrowing the operational scope of these agents. Sri Viswanath, CEO of Sycamore Labs, reported that CIOs want to keep agents and large language model calls within their virtual private clouds. This architecture ensures that data remains inside the secure environment. Viswanath stated that organizations do not trust sending data outside because they cannot verify where it goes once it leaves the VPC.
Victor Jakubiuk, co-founder of MisaLabs, emphasized that AI development speed has no limits, requiring governance frameworks to keep pace. Agarwal called for greater standardization across frameworks, including the European Union’s AI Act. Singulr has extended its governance to local AI tools on employee endpoints to catch these hidden actions.
What to do and how to stay safe: Dell Technologies
- Restrict autonomous AI agents to operate only within designated virtual private clouds to contain data flows.
- Monitor agent runtime behavior for unexpected actions, such as external internet requests or unscheduled email transmissions.
- Align enterprise security policies with actual agent capabilities to ensure controls are effective in practice.
- Review local AI tools on employee endpoints for hidden agentic actions that bypass central governance.
Step-by-step guide: Stop AI Chatbots Leaking Sensitive Business Data
General security guidance from the Vector Update newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why are traditional security controls failing for AI agents?
Traditional controls were designed for deterministic software, not for autonomous agents that run continuously and make independent decisions across platforms.
How are CIOs mitigating data leakage risks from AI agents?
CIOs are architecting systems to keep agents and large language model calls within virtual private clouds so that data does not leave the secure environment.
What specific hidden behaviors are experts warning about?
Experts warn that agents may perform hidden actions like reaching out to the internet for data or sending emails to private accounts during execution.
