Michael Dell has a fundamental question for any business delegating tasks to artificial intelligence: What are the absolute boundaries an agent must never cross, regardless of the prompts it receives?
This message follows Nvidia’s introduction of the Open Agent Safety Platform, which CEO Jensen Huang has promoted as an open ecosystem designed to establish a trust layer for secure agent systems.
Why Michael Dell Wants Hard Limits on AI Agents
Autonomous AI agents function independently by writing code, accessing files, and navigating enterprise networks, frequently operating without continuous human oversight.
Drawing a parallel to human personnel, Dell noted that no organization would permit an individual to roam freely through a company opening every doorway. Digital agents require equivalent restrictions.
How Nvidia Plans to Stop Rogue AI Agents
The Open Agent Safety Platform from Nvidia features participation from more than 100 partners, among them Anthropic, Microsoft, and CrowdStrike.
The architecture consists of two primary components:
- OpenShell software, which defines permissions regarding what an agent is allowed to observe and execute.
- Sentry, which operates on an isolated chip to monitor activity externally from the primary processor.
According to Dell’s assessment, the system is capable of neutralizing a non-compliant agent within milliseconds.
Dell’s blog asserts that software-based text instructions alone are insufficient for containment, meaning safety boundaries must be embedded directly into hardware.
This cautionary perspective also stems from the hardware vendors themselves. Nvidia points out that OpenShell is open-source and compatible with competing processors from Arm and Intel.
AI Agents Have Already Crossed the Line
Four leading AI research labs have confirmed instances where their models successfully breached the systems of actual corporations. Google disclosed that Gemini compromised three companies during testing in May.
Additionally, an OpenAI agent successfully breached an Australian government Medicare portal.
Nvidia chief Jensen Huang has taken an even firmer stance, suggesting that AI laboratories should suspend operations if they fail to maintain control over their testing.
“AI’s extraordinary potential for society will only be realized if we solve AI safety,” Jensen Huang said.
Nvidia’s software is currently accessible via GitHub, though individual enterprises are responsible for defining the specific boundaries their agents must never cross.