NVIDIA brings AI Agent security down to the chip level, delivering millisecond isolation for boundary breaches.

Beating AI Express News: NVIDIA has launched the Open Agent Safety Platform, composed of two components: OpenShell and Sentry. OpenShell sets permission boundaries for AI agents, while Sentry deploys monitoring and isolation on the BlueField-4 chip. If an agent crosses its boundaries, it can be isolated in milliseconds. OpenShell was open-sourced earlier, and can be directly applied to existing agents such as Claude Code and Codex without modifying the agents themselves. It restricts file access, network connections, tool calls, and credential usage, and enforces these rules during agent operation. Sentry is the newly added hardware security layer. It operates independently of agents, continuously monitoring agent behavior, tool calls, and data access, and handling threat detection and security policy enforcement. Even if an agent malfunctions, security control remains under the external hardware. Jensen Huang noted that over 100 industry partners are currently participating in the platform, including Anthropic, Microsoft, Salesforce, SAP, SpaceX, CrowdStrike, and Palantir. As agents start running continuously for hours or even days and interacting with code, tools, and enterprise data, NVIDIA is shifting security boundaries from within the model to the system and chip layers that the model cannot directly control.