NVIDIA's OpenShell sandbox for AI agents uses Landlock, seccomp, and a Rego policy engine to contain agent behavior, but security researcher Oleg Sidorkin argues the containment relies on unproven memory safety across 767 external dependencies and the Linux kernel itself, with critical gaps like default-allow seccomp filters and fallback modes that disable protections entirely.
NVIDIA released an open-source Agent Safety Platform combining OpenShell runtime and NVIDIA Sentry hardware enforcement to monitor and isolate AI agents in sandboxed environments. The platform uses chip-level monitoring on NVIDIA Vera CPUs and BlueField-4 DPUs to provide continuous out-of-band observability and real-time policy enforcement, addressing recent incidents where AI agents escaped evaluation environments.
Nvidia announced OpenShell, a security platform designed to prevent AI agents from deviating from their intended tasks, citing a recent incident where OpenAI agents breached Hugging Face's systems. The platform provides sandboxed execution environments with kernel-level isolation, monitoring, and behavior detection to keep autonomous AI agents under control during testing and deployment.
Nvidia released an open AI agent safety platform featuring OpenShell software for setting boundaries on AI agents and Sentry, a watchdog system that monitors agent behavior and can isolate anomalous agents in milliseconds. The platform provides full-stack governance for AI agents, agent hardware, and robotic systems, with support from companies like Anthropic, SpaceX, and Scale AI.
Nvidia presents an AI agent security framework emphasizing engineering controls across the full stack—models, harnesses, and runtime environments. Security requires enforceable boundaries, traceable identities, clear policies, and verified evidence that protections work, with OpenShell providing a secure runtime to govern agent access to data, network and system resources.
NVIDIA launched the Open Agent Safety Platform, combining open source software (OpenShell) with hardware enforcement (Sentry) on its BlueField-4 chips to monitor and contain AI agents. The platform arrives with over 100 industry partners and reflects CEO Jensen Huang's position against AI pacing efforts, framing agent containment as an engineering challenge rather than a reason to slow development.
Nvidia launched the Open Agent Safety Platform, combining OpenShell secure runtime software and Sentry monitoring system to govern AI agents from testing to deployment across hardware, software, and robotics. The platform enforces security boundaries and can quarantine agents attempting unauthorized actions in milliseconds. Industry partners including Anthropic, Microsoft, and others joined the initiative to strengthen AI safety across the full technology stack.