Nvidia introduced Openshell, an open software platform designed to set boundaries, monitor actions, and shut down artificial intelligence agents if they try to go rogue.
When was it introduced?
On September 28, 2026, the safety software platform was introduced through various digital channels, such as press releases through the Nvidia newsroom. Jensen Huang also made the live public announcement during a broadcast interview with CNBC and posted about it on his Twitter handle.
This security platform runs and oversees AI agents inside restricted and isolated sandboxes. It acts as the control plane managing API access, state, settings, and policy delivery. By combining NVIDIA OpenShell and NVIDIA Sentry, this two-layered security system stops an autonomous AI agent from going rogue.
How does this work?
• NVIDIA Open Agent Safety Platform works through three steps: Verify, Restrict, and Intercept. The system builds a zero-trust environment around the AI agent so that security lives entirely in the environment.
• Before an AI agent executes any code, YAML policy is fed into OpenShell’s Deterministic Policy Prover; it pre-analyzes all combined permissions to verify that no hidden access paths exist. It blocks the agent from starting if it catches risky or loophole-prone configurations.
• Once verified, the agent launches as a restricted child process on the host CPU inside the NVIDIA OpenShell runtime. If the AI agent tries to arbitrarily modify a local script or access unauthorized server data, Linux kernel controls instantly block the system and alert.
• NVIDIA Sentry steps in to handle cases, for example, when an AI agent tries to break out of the OS sandbox.
• Sentry captures raw telemetry data and checks whether an AI agent works against a strict, predefined behavioral profile.
Justin Boitano, Nvidia’s Vice President of Enterprise AI, has stated that Nvidia’s security platform would have prevented the Hugging Face security breach,
“From what we know, this new security platform could have stopped the breach if it was being used in frontier labs for model evaluation early on.”
He also emphasized the platform’s efficiency.
“OpenShell governs the agent’s actions, and then Sentry independently monitors and contains suspicious behavior… [Sentry] can quarantine a suspicious agent in milliseconds.”
Jensen Huang, through his X account, introduced Openshell and shared that AI is extraordinary technology and that people will believe in AI when the intention behind building it is safe and when it’s used with wisdom and responsibility. He emphasized on how safety and innovation should not be in conflict and trust will be earned through safety.