Nvidia Has The Right Idea On Advancing AI Agent Security: GuidePoint Exec

The approach Nvidia is taking with its new Open Agent Safety Platform is highly welcome, GuidePoint’s Victor Wieczorek tells CRN: ‘I think we need more of what Nvidia is doing.’

Nvidia’s newly announced Open Agent Safety Platform represents the type of engineering-focused effort needed to help the industry deploy and utilize AI agents in a more secure fashion, according to GuidePoint Security’s Victor Wieczorek.

Crucially, this is the case even if many businesses may not immediately benefit from the technology, said Wieczorek, senior vice president for offensive security at Herndon, Va.-based GuidePoint, No. 32 on CRN’s Solution Provider 500 for 2026.

Announced by Nvidia Monday, the Open Agent Safety Platform consists of software along with a hardware reference design aimed at providing stronger enforcement around the actions of AI agents—with the enforcement designed to occur independently of the agents themselves.

The launch of the Nvidia platform underscores the importance of bringing a greater level of outside expertise—as well as technical advancements—to the challenge of securing AI agents that can operate autonomously, according to Wieczorek.

“We need to see a lot more of those kinds of engineering ideas out in the world—because that’s how we’re going to get our way out of this,” he said in an interview Monday. “We’re going to build and solution our way out. We’re not going to finger point and just pontificate our way through.”

Ultimately, “I think we need more of what Nvidia is doing,” Wieczorek said.

The Nvidia launch comes amid growing concerns about the potential for AI agents to behave in unexpected ways, including through exceeding their intended access and permissions. A series of disclosures about such incidents have continued in recent weeks including from OpenAI, Anthropic and Meta.

In a blog post Monday, Nvidia pointed to the disclosures from frontier AI labs about agents escaping from their testing environments and, ultimately, accessing systems that were not authorized by the groups performing the tests.

Nvidia is looking to address those risks through measures including controls that live outside both the AI model and the software coordinating the model’s work, according to the company.

The platform includes the open-source OpenShell runtime software, which is used to isolate agents and restrict their access to resources, Nvidia said.

An additional component, Sentry, provides “in-silicon security enforcement,” the chip giant said in a news release. That means that if an AI agent tries to move outside its intended boundaries, Sentry “quarantines and stops it in milliseconds,” the company said.

Sentry runs on Nvidia’s BlueField-4 DPUs (data processing units), thus providing a separate mechanism for monitoring and enforcement within the hardware itself, according to Nvidia.

Certainly, the applicability of the Nvidia technology will depend to a large degree on how organizations are opting to utilize AI infrastructure, according to GuidePoint’s Wieczorek.

“There are, of course, limitations to how it’s applied and who can access it,” he said.

At present, the majority of businesses are likely not using Nvidia inferencing in the way that this type of technology could be applicable—“and it might be a long time before they are,” Wieczorek said.

Nevertheless, the announcement of Nvidia’s Open Agent Safety Platform is extremely valuable as a contribution toward how to solve the massive security challenges around AI agents, he said.

“We need more third-party voices at the table—both from an evaluator perspective and just a pure R&D one,” Wieczorek said.