Particle.news

Nvidia Unveils Open Agent Safety Platform to Stop Rogue AI Agents

The toolkit pairs open-source sandboxing with a proprietary hardware watchdog that enforces runtime limits and provides enterprise permissioning and audit trails.

Overview

  • Nvidia announced the Open Agent Safety Platform on Sept. 28, saying the package of OpenShell software and Sentry hardware would detect, constrain and shut down autonomous agents that break containment and that it could have stopped the July breakout that hit Hugging Face.
  • The platform combines OpenShell, an open-source secure runtime boundary, with Sentry, a proprietary monitor that runs on Nvidia BlueField-4 DPUs and can quarantine agents in milliseconds.
  • Nvidia says more than 100 companies are collaborating on or using the system, including Anthropic, Microsoft, SpaceXAI, JPMorgan Chase, Salesforce and SAP.
  • OpenAI is not a formal signatory even though it says it supports the work, and CEO Sam Altman has argued that hardware guardrails alone are incomplete and that broader oversight is needed.
  • Security researchers, enterprises and regulators are calling for independent audits, mandatory incident reporting and policy safeguards because Nvidia’s performance claims are self-reported and the platform’s hardware dependence raises questions about verification and vendor lock-in.