Dell AI Symposium highlights runtime governance challenges for autonomous agents
Executives warn that traditional controls fail for always-on AI agents, pushing governance to runtime.
Key points
- Autonomous agents run continuously across platforms, outpacing deterministic software controls.
- Singulr AI CEO Shiv Agarwal stresses aligning enterprise intent with actual runtime states.
- CIOs restrict agent operations to virtual private clouds to prevent external data leakage.
Enterprise leaders at Dell Technologies Inc.’s “AI Leadership Symposium” identified runtime AI governance as a critical operational challenge. As organizations shift from debating artificial intelligence adoption to active deployment, the complexity of managing autonomous agents has increased significantly. The discussion, covered by SiliconANGLE, highlighted that controls designed for deterministic software are inadequate for agents that operate around the clock across diverse platforms.
Shiv Agarwal, co-founder and chief executive officer of Singulr AI Inc., stated that rationalizing overlapping controls is difficult given the rapid pace of AI adoption. He emphasized the need to confirm that governance measures work in practice. According to Agarwal, someone must bridge the gap between enterprise intent and the actual runtime state of these systems.
Governance Complexity
Defining permitted actions for an agent is only one part of the governance problem. Enterprises must also assess behavior during execution, Agarwal explained. Singulr has extended its agent governance capabilities to local AI tools on employee endpoints to address this. He noted that risky actions, such as reaching out to the internet for data or sending emails to private accounts, can be deeply hidden inside agentic actions.
Victor Jakubiuk, co-founder of MisaLabs Inc., emphasized the need to keep pace with rapidly evolving AI capabilities. He stated that AI has no limit to the speed of development or progress. This velocity complicates efforts to maintain consistent oversight. Jakubiuk spoke with theCUBE Research’s John Furrier during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio.
Data Control Strategies
Some customers are responding to these governance challenges by narrowing where agents can operate. Sri Viswanath, founder and chief executive officer of Sycamore Labs Inc., reported that chief information officers want agents and large language model calls kept within their virtual private clouds. Viswanath explained that organizations are trying to architect systems so that all data stays within those boundaries.
Viswanath stated that CIOs do not trust sending data outside their controlled environments because they cannot track where it goes. This lack of trust drives the demand for strict containment strategies. The panelists discussed balancing standardization with rapid innovation during the symposium. Agarwal called for greater standardization across AI governance frameworks, including the European Union’s AI Act, to address these systemic issues.
What should you do?
Organizations deploying autonomous agents should review their runtime monitoring capabilities to ensure they can detect hidden actions such as external data requests. Security teams should verify that agent operations are confined to approved virtual private clouds to prevent unauthorized data exfiltration. Leaders should align their governance frameworks with specific enterprise intents to reduce the gap between policy and actual system behavior.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Dell Technologies
- Monitor runtime behavior of AI agents to detect hidden actions like external data requests or email sends.
- Restrict agent operations and large language model calls to internal virtual private clouds to contain data.
- Align governance frameworks with enterprise intent to ensure controls match actual system states.
- Review standardization efforts against frameworks like the EU AI Act to maintain consistent oversight.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why are traditional software controls insufficient for AI agents?
Traditional controls were designed for deterministic software, whereas autonomous agents run around the clock across a sprawl of platforms.
How are CIOs responding to data security concerns with AI agents?
CIOs are restricting agents and large language model calls to virtual private clouds to ensure data does not leave their control.
What specific hidden actions did Shiv Agarwal highlight as risks?
Agarwal highlighted actions such as reaching out to the internet to get data or sending emails to private accounts as hidden risks.
