NVIDIA's Open Agent Safety Platform launches with 100+ backers, including Anthropic and Microsoft but not OpenAI

NVIDIA's platform targets the problem that dominated this week: agents acting beyond their permissions. It has two layers. OpenShell is an open-source sandbox runtime that enforces kernel-level isolation on what an agent can touch. Sentry runs out-of-band on BlueField-4 DPUs, watching agent behavior from hardware the agent's host cannot tamper with and cutting off misbehaving agents within milliseconds. The platform covers the lifecycle from testing to production.

The partner list is the strategic story. More than 100 organizations signed on, including Anthropic, Microsoft, Oracle, CrowdStrike and Palantir. Claude Managed Agents already integrate for added governance. EDA leaders Cadence, Synopsys and Siemens are adopting it alongside Nemotron models in chip-design agents, and OpenClaw Enterprise lists OpenShell as an open governance option. Andrew Ng's OpenWorker harness is building on OpenShell, and he explicitly tied it to the OpenAI–Hugging Face breach. OpenAI's absence is being read in the community as a sign of diverging safety approaches.

Commercially, this is a classic NVIDIA move. Safety software is open-sourced, but the strongest guarantee, Sentry's out-of-band monitoring, requires BlueField hardware. Critics call that hardware lock-in. TechTarget quoted a sharper doubt: 'Nobody, including Nvidia, can yet show these platforms work.' The launch also came as NVIDIA added $150 billion to its buyback after a $96.22 billion quarter.

Watch for independent red-team results against Sentry, whether OpenAI joins or launches a rival standard, and whether cloud providers expose Sentry-style monitoring as a managed service.