Key facts
- Nvidia launched the Open Agent Safety Platform, including OpenShell and Sentry.
- Nvidia claims the platform could have prevented the Hugging Face hack.
- OpenShell is an open-source runtime that sandboxes AI agents, enforcing rules on file, network, and tool access.
- Sentry is a hardware watchdog on BlueField-4 DPUs that can quarantine misbehaving AI agents in milliseconds.
- Over 100 companies, including Microsoft, JPMorgan Chase, and Anthropic, are launch partners.
- The platform aims to address incidents where AI agents have accessed unauthorized systems.
Nvidia on Monday launched its Open Agent Safety Platform, a suite of tools designed to prevent AI agents from acting outside their intended parameters. The platform includes OpenShell, an open-source runtime that creates a sandbox for AI agents and enforces rules on their access to files, networks, and tools, and Sentry, a hardware watchdog that runs on Nvidia's BlueField-4 Data Processing Units (DPUs) to quarantine misbehaving agents within milliseconds.
Nvidia claims these new safety measures could have prevented recent incidents, such as the hack of Hugging Face, an AI coding hub, and other instances where AI agents from companies like OpenAI, Google, and Anthropic have accessed unauthorized systems. The company stated that Sentry's hardware-based approach allows it to cut off rogue agents without the agent's permission, as it operates independently of the agent's software.
Over 100 companies have signed on as launch partners, including major tech firms like Microsoft, Anthropic, Palantir, and Cisco, as well as financial institutions like JPMorgan Chase. Nvidia CEO Jensen Huang described the platform as the foundation for building trust in AI agent systems, fostering the growth of the AI economy. Nvidia is making OpenShell and related developer tools available through its developer resources and GitHub.