NVIDIA opens details of its agent safety platform
OpenShell confines agents to a sandbox with limited permissions, while Sentry monitors from a separate hardware layer and quarantines agents that cross boundaries in milliseconds; Anthropic, Microsoft and others have joined.
Jensen Huang just brought the entire tech cavalry together again to build a trust layer for the entire AI agent economy NVIDIA launched its Open Agent Safety Platform with 100+ organizations across AI, cybersecurity, infrastructure, robotics and finance Here’s the idea: Don’t trust AI agents to police themselves OpenShell puts the agent inside a secure sandbox with strict limits on what it can access and do Then Sentry watches from a separate hardware layer the agent can’t reach...and can quarantine it in milliseconds if it crosses those boundaries The safety controls live outside the AI itself This is AI safety being pushed all the way down into infrastructure and silicon SpaceXAI, Anthropic, Microsoft, Cisco, CrowdStrike, Palantir, Figure, Perplexity and many others are already working with it
This post quotes another post: x.com/jensenhuang/status/2104499465055023424