NVIDIA has launched an Open Agent Safety Platform designed to place enforceable controls around autonomous AI systems from testing through deployment.
The platform combines OpenShell, an open-source runtime that traces agent activity and enforces policies, with Sentry, a hardware-based watchdog that independently monitors agents and can quarantine them in milliseconds, according to NVIDIA. The company announced the platform on September 28.
NVIDIA said the platform is intended to address incidents in which AI agents have bypassed application-layer controls while carrying out assigned tasks. It combines software controls with monitoring at the hardware and compute layers rather than relying on the model or agent harness alone.
OpenShell Adds Runtime Controls for AI Agents
OpenShell provides a runtime security boundary for agents running on CPUs, controlling their access to systems, data and services while recording their actions. NVIDIA said the software is now broadly available as open source and can be extended to third-party compute platforms from Arm and Intel.
The runtime combines sandboxed execution, controlled service access and credential management with policy enforcement. A supervisor paired with each OpenShell sandbox checks outbound requests against policy, while the sandbox applies kernel-level filesystem and process controls to the agent workload.
Read more on agentic AI security: 65% of Enterprises Have Seen AI Agents Act Out of Scope
NVIDIA said OpenShell can also provide human oversight for agent activity. Salesforce, for example, has integrated it with Slack so users can review activity and audit events, and approve or reject requests for additional permissions.
More than 100 organizations are working with the Open Agent Safety Platform, including Anthropic, Microsoft, CrowdStrike, Palo Alto Networks, SAP, Salesforce and ServiceNow. NVIDIA said organizations in financial services, energy and robotics are also working with the technology, including for systems that can act in the physical world.
Sentry Adds Hardware-Based Agent Enforcement
Sentry is an optional hardware layer of the platform, using NVIDIA BlueField-4 data processing units (DPUs) as an out-of-band watchdog. NVIDIA said it continuously monitors agent behavior from an isolated trust domain and can quarantine or stop an agent that moves beyond its permitted boundaries in milliseconds.
The system is designed to operate independently of the agent itself, applying agent governance controls to requests involving data, tools, application programming interfaces (APIs) and services. NVIDIA said this provides a separate enforcement layer that remains outside the agent's software environment.
NVIDIA said Open Agent Safety Platform software, including OpenShell and related skills, is available through its developer resources and GitHub.
The launch comes amid growing concern over the behavior and permissions of autonomous systems. Guidance on agentic AI from the UK's National Cyber Security Centre (NCSC) previously warned that agents can have broad access to systems, data and tools and that rapid autonomous activity can make unexpected behavior harder to detect.
Image credit: credit: Sundry Photography / Shutterstock.com