NVIDIA Launches Open Agent Safety Platform: OpenShell Sandboxes Agents on Vera CPUs While Sentry on BlueFie…

What happened

NVIDIA launched the Open Agent Safety Platform, a new open-reference design aimed at enforcing AI agent safety externally rather than relying on the agents themselves. The platform uses OpenShell, an Apache 2.0 licensed runtime that sandboxes agents by applying policies defined in YAML files. It pairs this with Sentry, a watchdog running out-of-band on NVIDIA’s BlueField-4 data processing units (DPUs), which can identify and quarantine agents that break out of their sandbox in just milliseconds. Over 100 organizations are already working with this platform.

Why it matters

AI agents can behave unpredictably or maliciously, posing increased operational and security risks. NVIDIA’s approach separates the safety controls from the agent code, reducing the chance that a compromised or rogue agent can cause harm. Enforcing safety with a runtime sandbox and a DPU-level quarantine means organizations can run complex AI workflows with tighter security and real-time threat containment. This changes the default risk profile for running autonomous AI, especially in sensitive environments where failure or bad behavior could be costly or dangerous.

What to watch next

The real test will be how widely the platform is adopted beyond those early 100 organizations and how well OpenShell policies cover diverse real-world AI workloads. Integration with existing AI ops and monitoring tools will be crucial for operational teams. Watch for new partnerships with cloud providers or enterprise software vendors that could embed this safety platform into broader AI infrastructure. Also track any benchmarks or case studies showing Sentry’s quarantine effectiveness under live workloads for high-risk or mission-critical AI deployments.

AI Quick Briefs Editorial Desk