NVIDIA launches Open Agent Safety Platform to contain AI agents
NVIDIA's Open Agent Safety Platform adds runtime controls, OpenShell sandboxing and Sentry monitoring for AI agents.
NVIDIA has introduced the Open Agent Safety Platform, an open reference design meant to keep AI agents inside defined technical boundaries as they move from experiments into production systems. The company says the platform combines OpenShell, its open source runtime for agent governance, with Sentry, a hardware assisted security layer that can monitor and enforce policy outside the agent environment.
The practical shift is where control happens. Model safeguards and prompt rules influence what an agent tries to do. NVIDIA is arguing for runtime controls that decide what the agent is allowed to access, change, or call after it starts taking actions. That matters for organizations testing agents that can use tools, run code, query internal systems, or request credentials.
What NVIDIA is shipping
On NVIDIA's platform page, OpenShell is described as a sandboxed execution and policy enforcement layer for governing what agents can see and do. Sentry adds out-of-band monitoring using NVIDIA DOCA and BlueField hardware, with NVIDIA saying the system can provide telemetry, identity governance, and policy enforcement from a separate security domain. The company also says Sentry can quarantine agents in milliseconds when they move outside approved boundaries.
NVIDIA lists Vera CPU and BlueField DPU systems as optimized targets, but says the broader platform is compatible with other hardware. The OpenShell FAQ says it can run without BlueField-4 on supported local, on-premises, cloud, and Kubernetes infrastructure. That distinction is important for buyers because the open runtime and the hardware-isolated monitoring layer are not the same deployment decision.
Why teams should care
Independent coverage from WIRED reported that OpenShell is moving into general release and that NVIDIA is positioning the broader platform around recent concerns about autonomous agents escaping containment or probing systems they should not touch. Constellation Research reported that NVIDIA described the effort as a full-stack governance and control system for agents, with partners across enterprise software, cloud, cybersecurity, and AI infrastructure.
The useful decision rule is simple: treat agent security as an infrastructure problem, not only a model behavior problem. Teams running agents with access to source code, customer data, SaaS tools, or production systems should look for enforceable allow and deny decisions, audit trails, and a way to revoke permissions while work is in progress. NVIDIA's approach will still need real-world proof across mixed hardware and cloud estates, but it puts a concrete technical boundary around a risk that many agent pilots currently handle with trust and monitoring after the fact.
Sources
Cover photo by Santhosh Kanthala on Pexels, used under the Pexels License.
Comments (0)
Leave a Comment