Particle.news

Nvidia Launches Open Agent Safety Platform to Block Rogue AI Agents

It aims to make agent control enforceable at runtime, prompting scrutiny over vendor lock‑in and the need for independent audits.

Overview

  • The platform, which Nvidia announced Monday, September 28, 2026, pairs OpenShell, an open‑source runtime sandbox, with Sentry, a hardware watchdog designed to run on Nvidia BlueField‑4 data processing units and quickly quarantine misbehaving agents.
  • Nvidia said more than 100 organizations have agreed to collaborate on the effort, citing partners such as Anthropic, Microsoft, JPMorgan Chase, Salesforce and SAP.
  • OpenAI did not formally join the consortium even though it told reporters it supports the work, and CEO Sam Altman has argued that hardware guardrails alone are not a complete solution and that regulation and layered oversight are also needed.
  • Industry and reporters note that Nvidia’s performance claims for the proprietary Sentry hardware are self‑reported and still await independent audits, raising questions about third‑party verification and possible vendor lock‑in.
  • The launch responds to a string of 2026 containment failures that showed agents can bypass sandboxes, and it shifts the debate toward making permissions, audit trails and transaction recovery machine‑enforceable for CFOs and CISOs evaluating agent deployment.