NVIDIA introduces the Open Agent Safety Platform, an open software platform and reference design for continuous monitoring and enforcement controls over autonomous AI agents from testing through production deployment. The platform integrates NVIDIA’s OpenShell secure runtime with NVIDIA Sentry, an out-of-band hardware watchdog designed to halt agents that exceed their designated boundaries. This launch addresses a key security challenge for agentic AI, which differs from conventional chatbots in its ability to use tools, access files and databases, call APIs, execute code, and potentially operate for extended periods with delegated credentials.
OpenShell serves as the software foundation. Drift can arise from ambiguous instructions, failed tool calls, missing capabilities, code bugs, or long-running tasks where the agent frequently experiments with alternative approaches.
Sentry is described as an out-of-band watchdog that independently observes agent activity, enforces policies in silicon, and can quarantine or stop an agent within milliseconds when it attempts to leave its permitted software boundary. The company claims this placement enables real-time monitoring and policy enforcement at line speed, while keeping the security mechanism invisible to agents and potential attackers. By integrating sandboxing, verifiable policy, identity checks, behavioral telemetry, and hardware-enforced intervention, NVIDIA aims to establish a security boundary for autonomous AI similar to the sandboxing model that helped make web applications safer to deploy at scale.
Explore for your team.
