NVIDIA released the Open Agent Safety Platform, including a tool called OpenShell, to secure AI agents at the sandbox and hardware level, according to a CyberScoop report. 

The company announced it September 28, amid growing concern after several AI models escaped sandbox protections during testing and raised alarms about real-world breaches.

OpenShell, built on open-source software, lets AI system operators define and test an agent’s permissions, files, networks, tools, before deploying it into enterprise networks. NVIDIA is also adding security features to its BlueField 4 chip to monitor agent behavior at the hardware level, according to the report.

More than 100 organizations, including Anthropic, Microsoft, Palantir, and JPMorgan Chase, are backing the initiative and plan to integrate or support the platform, according to the report. 

Can AI Agents Be Safely Contained?

After models from OpenAI, Anthropic, and Meta escaped sandbox protections during testing and breached real organizations this year, some AI industry leaders, including Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman, suggested AI systems may now be too advanced to fully contain, according to CyberScoop.

NVIDIA CEO Jensen Huang has argued that frontier AI companies and their supply chain partners need to improve their own security posture, rather than treat containment as a losing battle.

Huang: Agent Security Is “an Engineering Problem”

In a September 28 interview with CNBC, Huang said he sees agent security as solvable. He said, “I believe it’s an engineering problem, I know it’s an engineering problem, and we all need to hope that it’s an engineering problem.”

Huang has also said he supports government regulation of the AI industry only if it is designed to promote growth and innovation, a position he reiterated last week, according to the report.

“Trust Is How Safety Is Earned,” Huang Says

On X, Huang called the new platform “the beginning of an open ecosystem to build the trust layer for safe agent systems.” He said security has to become “foundational” to AI development.

“Trust and innovation are not in conflict,” Huang wrote. “Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI.”

Why Security Now Has to Sit Outside the Model Itself

Aviv Nahum, CEO of Above Security, an AI-native insider risk management platform that investigates both human employees and autonomous AI agents, told CyberScoop that NVIDIA is essentially arguing that some enforcement has to live outside the model itself, in a layer the agent can’t reason around or modify. 

“What this announcement says to me is that the industry is finally converging on a basic cybersecurity principle: model alignment is not a substitute for security engineering,” Nahum said. “Sandboxes, identity, least privilege, independent monitoring, and containment are not new ideas. What is new is that we now have autonomous software capable enough that failing to apply those principles becomes much more consequential.”

Outside Enforcement Is the Same Principle CDOs Already Apply Elsewhere

For chief data officers (CDOs), Nahum’s point is clear: safety controls can’t live only inside the AI system they’re meant to constrain. 

An AI agent that can reason around its own guardrails needs enforcement it can’t touch or change, similar to how a person who requests access to a system shouldn’t be the same person who approves it.

Only time will tell whether NVIDIA’s platform becomes a standard. With more than 100 companies backing it, “enforcement outside the model” looks like the industry’s practical answer to a problem that policy alone hasn’t solved. 

For AI governance, that gives CDOs a clear question to ask vendors: “What independent layer stops the AI agent if its training or internal guardrails fail?”