NVIDIA launched the Open Agent Safety Platform to contain AI agents and prevent security breaches, following incidents like the Hugging Face incident involving over 17,000 agents. The system includes OpenShell and Sentry, designed to monitor and enforce policy limits.
NVIDIA’s Open Agent Safety Platform aims to contain AI agents and prevent security breaches by enforcing strict behavioral limits. The platform emerged after recent incidents, including the Hugging Face incident, where over 17,000 agents attacked infrastructure, according to CNBC. NVIDIA CEO Jensen Huang described the tool as a browser for agents,
emphasizing the need to confine AI operations to prevent uncontrolled behavior.
Core Components: OpenShell and Sentry
The platform comprises two key tools: OpenShell and Sentry. OpenShell, an open-source secure runtime, runs on central processors and sets formal limits on agent capabilities, ensuring they access only what is necessary for their tasks. Sentry monitors agents and runs on network chips. According to Nvidia, Sentry operates independently of the agent, providing observability and enforcing security policies.
OpenShell’s policy enforcement is rooted in zero-trust principles, requiring agents to operate within verifiable constraints. The system checks access limits—files, networks, tools, processes, and credentials—before execution and enforces them during operation. Nvidia notes that this approach addresses drift, where agents depart from the intended task or operating constraints due to a policy block, a bug, a missing tool, ambiguous instructions, or running for days or weeks.
Hardware Integration and Industry Partnerships
The platform is optimized to run on NVIDIA Vera CPU- and BlueField DPU-based systems, and is also compatible with other hardware systems, with Sentry extending monitoring and enforcement into NVIDIA BlueField hardware. This integration allows real-time, out-of-band observability, ensuring security even if host resources cannot be trusted.
NVIDIA is also working with Anthropic to integrate cloud managed agents with OpenShell. The company emphasized that the platform is a reference design, encouraging third-party development to expand its reach across diverse hardware ecosystems.
Addressing Recent Security Incidents
Justin Boitano, vice president of enterprise AI at Nvidia, stated the platform could have prevented such incidents, citing that Hugging Face reported over 17,000 agents attacking their infrastructure that went on for days and weeks.
NVIDIA’s approach addresses this by controlling the path to the model,
ensuring security policies are enforced at the infrastructure level. The company’s sources emphasize that this architecture allows for resilient security, with continuous evaluation of the security and integrity of agents against a predefined behavioral profile.
Future Implications and Industry Adoption
NVIDIA’s platform signals a shift toward AI safety, with the potential to shape the agent economy similarly to how the internet evolved with trust. Industry leaders and partners are already supporting the platform.
The platform is optimized to run on NVIDIA Vera CPU- and BlueField DPU-based systems, and is also compatible with other hardware systems. With organizations adopting the system, the focus now shifts to widespread implementation. As AI agents become more prevalent, the success of this platform could set a precedent for industry-wide safety standards.