NVIDIA launches an open AI agent safety platform with over 100 partners
NVIDIA's Open Agent Safety Platform combines the OpenShell secure runtime with Sentry hardware monitoring to govern AI agents from testing through deployment.
NVIDIA has introduced its Open Agent Safety Platform with support from more than 100 organisations. The open initiative is designed to place enforceable controls around AI agents from testing through deployment. According to NVIDIA, its two principal components are OpenShell and Sentry.
What OpenShell does
OpenShell is an Apache 2.0-licensed open-source runtime that executes agents in an isolated environment. Administrators can define limits on network access, files, processes, tools, and credentials before a workload starts. The controls sit outside the model and agent harness, so unexpected agent behaviour does not automatically translate into unrestricted access.
NVIDIA says OpenShell is optimised for its Vera CPU but can also be extended to third-party compute platforms, including systems based on Arm and Intel technology.
Sentry's hardware layer
Sentry is an out-of-band watchdog designed to run on BlueField-4 data processing units. In the Vera Rubin POD reference design, communications travel through a path that this layer can monitor independently of the host operating system. Using NVIDIA DOCA, Sentry correlates agent activity, policy decisions, and access to tools and data. NVIDIA claims it can quarantine an agent within milliseconds if the agent moves beyond its defined boundaries.
Why the architecture matters
The design separates the application, runtime, and physical infrastructure layers. OpenShell policies determine what an agent is allowed to do, while Sentry provides an independent observation and enforcement layer. This architecture may limit the impact of mistakes or abuse, but the platform announcement alone does not demonstrate that every agent or escape path will be contained in practice; effectiveness will depend on configuration, implementation, and independent evaluation.
NVIDIA says organisations already using Vera and BlueField-4 systems will be able to activate the protections through a software update. Companies spanning models, cybersecurity, infrastructure, and enterprise software are participating in the wider ecosystem.

Source: AI News