NVIDIA has launched an open software platform designed to give organizations greater control over artificial intelligence agents and limit what they can access and do.
The NVIDIA Open Agent Safety Platform combines software controls with a hardware security reference design. The company announced the platform Monday, saying it establishes boundaries across AI software, computing infrastructure and robotics.
The launch follows disclosures about AI agents acting outside their intended limits. A Hugging Face security breach was driven primarily by an internal research model, OpenAI said in its account of the incident.
As previously reported by The Dallas Express, NVIDIA announced plans in 2025 to build AI supercomputers in Texas with manufacturing partners.
OpenShell establishes agent boundaries
A central component of the platform is NVIDIA OpenShell, open-source software that creates an isolated environment for AI agents.
OpenShell allows operators to establish policies governing an agent’s access to files, tools, networks, processes and credentials. The software tracks agent activity and enforces those policies while an agent performs tasks, according to NVIDIA’s technical description.
OpenShell runs on NVIDIA’s Vera CPUs, and developers can extend it to work with third-party computing platforms, including those from Arm and Intel, the company said.
Justin Boitano, NVIDIA’s vice president of enterprise AI, said the system could have prevented the Hugging Face breach.
“From what we know, this new security platform could have stopped the breach if it was being used in frontier labs for model evaluation early on,” Boitano said, CBS News reported.
Sentry adds hardware-based monitoring
The platform’s reference design includes NVIDIA Sentry, a separate security layer designed to continuously monitor AI agents.
Sentry runs on NVIDIA BlueField-4 data processing units and operates independently from the agent. It can detect attempts to move beyond established boundaries and quarantine an agent within milliseconds, NVIDIA said.
The system uses NVIDIA DOCA software to inspect agent requests and responses, verify agent identity and enforce access policies for data, tools, APIs and services, according to the company.
The design separates security controls from the AI agent to make them harder for agents or attackers to bypass, NVIDIA said.
More than 100 organizations involved
More than 100 organizations are working with the platform’s technologies, including Anthropic, Cisco, CrowdStrike, Dell Technologies, JPMorganChase, Microsoft, Palantir, Palo Alto Networks, Salesforce, SAP, Scale AI and ServiceNow, NVIDIA said.
Anthropic’s Claude Managed Agents use integrations with OpenShell and BlueField to add controls. Salesforce has integrated OpenShell with Slack so users can review agent activity and audit events and approve or reject requests for additional permissions, according to NVIDIA.
Scale AI is incorporating the technology into its agent infrastructure, while robotics companies including Figure, Gecko Robotics and Skild AI are using OpenShell to add security controls to autonomous systems, the announcement said.
Financial services companies including Citi and JPMorganChase are also collaborating with NVIDIA on open-source agent safety technology, the company said.
OpenShell available as open source
OpenShell is available through NVIDIA’s developer resources and GitHub under the Apache 2.0 license.
The company said the platform is intended to support cooperation on AI safety through shared tools, research and security practices. Its work also supports the Open Secure AI Alliance, a Linux Foundation-governed initiative with more than 120 organizations.
AI safety requires controls across the full technology stack, NVIDIA CEO Jensen Huang said in the announcement.
Provided by Dallas Express
