NVIDIA has launched a new Open Agent Safety Platform aimed at giving companies stronger control over AI agents as these systems increasingly perform tasks with limited human intervention.

The platform, announced on September 28, includes OpenShell, open-source software that creates security boundaries around AI agents, and Sentry, a reference system designed to continuously monitor agent activity and quarantine an agent if it moves outside approved limits.

NVIDIA said Sentry operates through its BlueField-4 data processing units, while OpenShell provides runtime controls for agents operating on CPUs. The company says the technology can enforce policies governing access to data, applications, tools and other services. OpenShell is also designed to work with computing platforms from companies including Arm and Intel.

The announcement comes as the technology industry faces growing attention over AI agents that can independently browse systems, use software and perform complex tasks. Recent incidents involving AI systems operating beyond intended boundaries have increased interest in security controls that work outside the AI model itself.

NVIDIA said more than 100 organizations are working with its platform or related technologies, including Microsoft, Anthropic, Cisco, CrowdStrike, JPMorganChase, Salesforce and SAP. The company is positioning the system as an open approach to AI-agent security across enterprise software, computing infrastructure and robotics.

The development highlights a growing technology focus in the U.S.: building security and governance systems alongside increasingly autonomous AI tools, rather than relying only on controls inside individual AI models.