Nvidia launches Open Agent Safety Platform for AI-agent containment
A post describes OpenShell as a software sandbox that limits agents’ access to files and passwords, paired with a separate Sentry safeguard on BlueField-4 hardware.
TLDR
Posts describe Nvidia’s platform as combining an OpenShell software sandbox with hardware monitoring on a separate BlueField-4 processor. One says Sentry can cut off an agent’s connection to its language model if it detects an escape attempt; another says the hardware can quarantine agents in milliseconds if they try to access unauthorized systems.
Combined views
—
2 Sources, first seen 1h ago
