[NVIDIA Brings AI Agent Security Down to the Chip Level: Millisecond-Scale Isolation for Boundary Breaches] Beating AI Newsflash: NVIDIA has unveiled the Open Agent Safety Platform, consisting of two components—OpenShell and Sentry. OpenShell defines permission boundaries for AI Agents, while Sentry shifts monitoring and isolation to the BlueField-4 chip level. If an Agent crosses its boundaries, it can be isolated within milliseconds. OpenShell has already been open-sourced and can be directly applied to existing Agents like Claude Code and Codex without modifying the Agents themselves. It is responsible for restricting file access, network connections, tool invocation, and credential usage, enforcing these rules during Agent runtime. Sentry, the newly added hardware security layer, operates independently of the Agent, continuously monitoring the Agent's behavior, tool usage, and data access. It handles threat detection and security policy enforcement. Even if the Agent itself malfunctions, security controls are still managed externally by hardware. Jensen Huang stated that over 100 industry partners are already participating in this platform, including Anthropic, Microsoft, Salesforce, SAP, SpaceX, CrowdStrike, and Palantir. As Agents begin running continuously for hours or even days, interacting with code, tools, and enterprise data, NVIDIA is shifting security boundaries from within the model to system and chip layers that the model cannot directly control.