AI · 6h ago
NVIDIA on September 28 launched its Open Agent Safety Platform, pairing the OpenShell runtime with Sentry, a watchdog built on BlueField-4 data processing units, to constrain autonomous AI agents from testing through deployment. NVIDIA says more than 100 organizations are working with the platform.
OpenShell sandboxes agents and enforces rules on files, processes, networks, tools, and API access. Sentry runs from separate hardware and watches from an out-of-band trust domain, so it can quarantine or stop a misbehaving agent in milliseconds instead of relying on the agent or its harness to police itself.
For teams giving agents real credentials or long-lived access to internal systems, the shift is where enforcement lives: outside the agent, on infrastructure that still sees the requests when the agent drifts. If an assistant can reach services, data, or APIs on your behalf, that control plane now matters as much as the model layer.
5 sources covering this story
As AI world debates security, NVIDIA releases open source tools for agents
NVIDIA launches the Open Agent Safety Platform and OpenShell to secure AI runtime environments, preventing rogue AI agent breaches with hardware and software controls.
NVIDIA Launches Open Platform to Secure Autonomous AI Agents
NVIDIA has launched a platform pairing runtime controls with hardware monitoring for AI agents
NVIDIA wants AI agent safety enforced in silicon, not left to the agent - Help Net Security
NVIDIA's Open Agent Safety Platform pairs OpenShell runtime controls with hardware monitoring to keep AI agents within set limits.
Nvidia Unveils AI Agent Safety Platform With Hardware-Based Watchdog
The platform combines open source software and a reference system design to keep AI agents within set boundaries.
CrowdStrike and NVIDIA Extend Security Across the AI Stack
CrowdStrike and NVIDIA are collaborating on the NVIDIA Open Agent Safety Platform, an open reference design to extend security deeper into the agentic stack and establish stronger boundaries for autonomous AI.
Part of the PlainSec briefing for 2026-09-28