NVIDIA introduces open platform to control AI agents

NVIDIA has launched a new open platform that enables the secure deployment of AI agents and prevents them from going out of control, called the NVIDIA Open Agent Safety Platform. The platform aims to impose clear restrictions on the agents’ permissions, continuously monitor their behavior, and intervene to isolate them within fractions of a second if they attempt to exceed their authorized capabilities. It includes an open-source secure operating system (OpenShell) and the NVIDIA Sentry technology, which operates at the hardware level to protect agents independently, relying on BlueField-4 data processing units. NVIDIA collaborated with over 100 companies and institutions, including Microsoft, Cisco, and SpaceXAI, to monitor instances where agents bypass testing protocols and to recognize that traditional methods of safeguarding AI models are no longer sufficient as agents gain the ability to execute external tools and services autonomously. The platform provides an isolated environment that enforces rules before and during the agent’s operation, with an immutable decision log, and employs in-hardware protection mechanisms (In-Silicon Security) to ensure agents do not violate security boundaries. NVIDIA believes that assessing agent trustworthiness requires a multi-layered security infrastructure, encompassing the applications, runtime environment, and underlying infrastructure, supported by Vera processors, BlueField units, and integration with systems from other companies. This platform is considered a significant step towards comprehensive and reliable control of intelligent agents, especially as reliance on tools capable of handling data and executing commands independently continues to grow.

articleDetails.fontSize.sizeحجم النص

تنويه: هذا ملخص تم إنشاؤه بواسطة الذكاء الاصطناعي

عرض المقال بالكامل