AI agents are being designed to do more than answer questions. They can access files, use workplace tools and carry out tasks with less human supervision. But giving software that much freedom creates a new risk. Nvidia Corp (NVDA) wants to make these digital assistants safer to use, though BNP Paribas analyst Karl Ackerman sees a potential trade-off: runtime could slow by as much as 25%.

Giving AI More Freedom, With Limits

Nvidia announced its OpenShell reference architecture at GTC Berlin recently, building on open-source software designed to control what AI agents can access and do. The system is intended to set boundaries around agents as they work, monitor their actions and block activity that violates security rules.

That matters because an AI agent does not simply produce an answer and stop. It may access company files, connect to online services, run code or use other tools to complete a task. If it misinterprets an instruction or is manipulated, those permissions can create risks beyond a misleading chatbot response.

Nvidia's approach puts security controls outside the AI agent itself, rather than relying on the agent to police its own behavior. Its reference architecture combines computing systems for running AI with BlueField-4 data processing units and Nvidia's Sentry security framework to monitor activity and enforce restrictions.

Companies including Cisco Systems, Inc. (CSCO), CrowdStrike Holdings, Inc. (CRWD), Microsoft Corp (MSFT) and Alphabet Inc's (GOOGL) (GOOG) Google Cloud are working with the architecture, Nvidia said.

The Potential Cost of Safety

In his note following GTC Berlin, Ackerman said OpenShell could slow runtime by as much as 25%. That is an analyst estimate, not a confirmed performance result across all tasks or deployments.

The potential trade-off is straightforward: More checks can mean more work before an AI agent is allowed to act. For businesses, the question is whether added safeguards justify any delay, particularly when agents are expected to handle time-sensitive work or perform thousands of tasks at once.

Still, a slower but controlled agent may be more useful to an enterprise than a faster system that can access information or take actions it should not. The actual balance will depend on the task, security settings and implementation.

What It Means for Nvidia

OpenShell reflects Nvidia's effort to sell more than the chips that power AI. As companies move from testing AI assistants to giving them access to real business systems, security could become an increasingly important part of enterprise AI spending.

For investors, the key question is whether businesses adopt Nvidia's approach at scale and whether its security controls can protect agents without materially hurting performance. Independent benchmarks and customer deployments will matter more than the launch announcement alone.

If Nvidia can make AI agents safer without making them meaningfully slower, security could become another reason businesses choose its broader platform.