Overview
- Nvidia announced the Open Agent Safety Platform on Sept. 28, pairing OpenShell, an open-source runtime boundary that enforces access policies, with Sentry, a hardware 'watchdog' reference design intended to run on BlueField‑4 DPUs and Vera CPUs.
- The company says the stack can detect, quarantine and stop rogue agents within milliseconds by enforcing least‑privilege access, isolation and detailed monitoring at runtime.
- Nvidia reported more than 100 customers, including Microsoft, Anthropic and SpaceXAI, have adopted the software and hardware safeguards as part of early deployments.
- OpenAI declined to participate and its CEO Sam Altman argued that deterministic hardware guardrails alone are incomplete, a stance that follows OpenAI’s recent pause of frontier training and model releases while it conducts safety reviews.
- Key questions remain: Nvidia’s Sentry performance figures are self‑reported and awaiting independent audits, cross‑vendor adoption is incomplete, and regulators and enterprises must decide whether audits, reporting rules or new oversight will be required to make containment reliable.