NVIDIA Launches Open Agent Safety Platform: OpenShell Sandboxes Agents on Vera CPUs While Sentry on BlueField-4 Quarantines Them in Milliseconds
NVIDIA just shipped the hardware kill switch for rogue agents. OpenShell sandboxes them on Vera; Sentry on BlueField-4 quarantines in milliseconds. Over 100 orgs are already using it.
Why it matters
NVIDIA has released an open reference design for agent safety that enforces containment outside the agent itself—using policy-driven sandboxing (OpenShell on Vera CPUs) and hardware-level watchdog quarantine (Sentry on BlueField-4 DPUs). This addresses a material risk in production agent deployments: an agent that escapes its intended boundary can be stopped by dedicated hardware in sub-millisecond time, independent of the agent's own controls. Early adoption signals (100+ organizations) suggest this is moving from concept to operational infrastructure.
The key facts
5 to knowOpenShell: Apache 2.0 runtime sandbox enforcing YAML policies on Vera CPUs
Sentry: out-of-band watchdog on BlueField-4 DPUs, quarantine latency in milliseconds
100+ organizations working with the platform
Open reference design (not proprietary NVIDIA-only stack)
Enforcement outside the agent itself (hardware/runtime layer, not agent-internal guardrails)
The story so far
Earlier coverage of this storyline
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: NVIDIA has launched the Open Agent Safety Platform, an open reference design that enforces AI agent safety outside the agent itself. OpenShell, an Apache 2.0 runtime, sandboxes agents under YAML policies. Sentry, an out-of-band watchdog on BlueField-4 DPUs, can quarantine an agent that escapes its…