Chip giant Nvidia CEO Jensen Huang unveiled a new security architecture at the GTC conference in Berlin, which has an Israeli development core. The technology is designed to address a key challenge of the AI era: how to prevent AI agents, software capable of performing tasks autonomously, from acting against instructions and putting organizational data at risk.
As companies grant these agents access to internal systems and sensitive documents, as well as the ability to perform actions on behalf of employees, the risks of errors, data leaks, and the misuse of granted permissions increase. Nvidia’s solution enables organizations to predefine the scope of an agent’s permitted actions, monitor its activity, and automatically halt it should it attempt to carry out a prohibited operation.
A defense layer based on a chip developed in Israel
The new platform is based on OpenShell, an open-source software previously introduced by Nvidia, which is now integrated with hardware-level security mechanisms. At the heart of this hardware-based defense layer lies the BlueField-4 chip. Developed over the past few years at Nvidia’s Israel R&D center, the chip utilizes technology that originated with Israeli company Mellanox, which Nvidia acquired in 2020.
Running on top of this chip is a separate system called Sentry, which monitors agent activity and independently enforces security policies. This means that supervision does not rely solely on the software running the agent. Nvidia says that the system can detect anomalies and isolate an agent within milliseconds. The company is also promoting this solution in collaboration with Claude developer Anthropic.
In his conference address, Huang explained that the goal is to enable companies worldwide to adopt AI agents and deploy them autonomously within their organizations, with access permissions tailored to each specific task and user. He noted that every action involving an agent’s capabilities, the assigned task, and the identity of the employee operating it requires a distinct level of permission. Therefore, unlike systems where permissions are predefined and static, supervising AI agents must be dynamic, with security policies adjusted every time an agent is activated.
Huang said, "Every time an agent is activated, the permissions are different. Therefore, permission management, policy enforcement, and monitoring must be carried out dynamically." Huang, described OpenShell as a revolutionary technology.
For the Israeli cybersecurity industry, this represents a major development in a sector poised for growth as the use of AI agents within organizations expands. But it also presents a competitive challenge. Nvidia aims to make security an integral part of the AI infrastructure itself, rather than merely an external product added on.
Full disclosure: the journalist was a guest of Nvidia at the GTC Conference in Berlin.
Published by Globes, Israel business news - en.globes.co.il - on October 21, 2026.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2026.