NVIDIA Launches AI Agent Safety Platform With Israeli Contributions

NVIDIA has announced the NVIDIA Open Agent Safety Platform, a new system designed to enhance the security of AI agents from testing through deployment. The platform integrates open-source software and reference architecture to enforce operational policies and controls across software, hardware, computing, and robotic systems.

A significant portion of the platform was developed in Israel, positioning NVIDIA to compete with numerous Israeli cybersecurity firms offering similar AI agent security solutions. Several Israeli cybersecurity companies, including Argyle, Torq, and Upwind, are among those partnering with NVIDIA on this initiative.

The platform features Centaur, a key Israeli development operating on BlueField-4 DPUs, also developed at NVIDIA's Israeli research center. Centaur continuously monitors AI agent activity and independently enforces security policies at the chip level, incorporating threat detection, agent control, policy enforcement via hardware, and data access protection within a secure, isolated environment.

NVIDIA CEO Jensen Huang emphasized the critical need for AI safety to realize the technology's potential, stating, "AI safety requires engineering work across all layers of the system." He highlighted the platform's role in fostering collaboration among industry, researchers, and public sector organizations to share best practices and advance international cooperation on AI safety standards.

NVIDIA also detailed collaborations with companies like Anthropic, which is integrating security and control layers into its Claude Managed Agents. Salesforce, SAP, CrowdStrike, and Cisco are also involved. SpaceXAI is utilizing the platform for its Cursor programming agents and Grok models, with President Mike Nichols noting the importance of external controls that agents cannot bypass to ensure customer-defined boundaries are maintained for Cursor and Grok.