Nvidia embeds AI Agent security down to the chip level: out-of-bounds can be isolated in milliseconds.

动察 Beating AI News Flash: NVIDIA has launched the Open Agent Safety Platform, consisting of two parts: OpenShell and Sentry. OpenShell sets permission boundaries for AI Agents, while Sentry places monitoring and isolation on the BlueField-4 chip side. Once an Agent crosses the line, it can be isolated within milliseconds.

OpenShell was previously open-sourced and can be directly wrapped around existing Agents such as Claude Code and Codex without modifying the Agent itself. It is responsible for restricting file access, network connections, tool calls, and credential use, and enforces these rules while the Agent is running.

Sentry is the newly added hardware security layer this time. It runs independently of the Agent, continuously monitors the Agent's behavior, tool calls, and data access, and is responsible for threat detection and security policy enforcement. Even if the Agent itself behaves abnormally, security control is still carried out by external hardware.

Jensen Huang stated that more than 100 industry partners have already participated in this platform, including Anthropic, Microsoft, Salesforce, SAP, SpaceX, CrowdStrike, and Palantir. As Agents begin to run continuously for hours or even days and come into contact with code, tools, and enterprise data, NVIDIA is starting to move the security boundary from inside the model to the system and chip layers that the model cannot directly control.