securityweek.com
Sep 28, 2026
Nvidia Unveils AI Agent Safety Platform With Hardware-Based Watchdog
AI-summarised brief · reviewed before publication
Nvidia introduced an AI agent safety platform that couples software sandboxing with a hardware-based watchdog called Sentry. The platform, built on OpenShell, manages sandbox lifecycles, enforces kernel-level controls, and logs policy decisions. Network traffic from agents is inspected by a supervisor that can allow reads while blocking writes, and API keys are substituted outside the agent’s environment. Sentry, running on BlueField‑4 DPUs, provides in‑silicon enforcement, quarantining agents that breach boundaries within milliseconds. Over 100 organizations, including Anthropic, Salesforce, and SAP, are testing the solution.
💡 Why It Matters
- · By embedding hardware enforcement, Nvidia closes a critical gap left by software‑only safeguards, ensuring that even compromised hosts cannot bypass AI agent controls.
- · This dual‑layer approach sets a new standard for secure AI deployment in enterprise environments.
