NVIDIA Launches Open Agent Safety Platform for Full-Stack AI Security

NVIDIA has introduced the Open Agent Safety Platform, an open software platform and reference architecture intended to secure autonomous AI systems from testing through deployment.

The initiative is designed to connect controls across the agent runtime, underlying compute infrastructure and physical systems that execute agent-directed tasks. More than 100 companies are participating, according to NVIDIA and launch partner SentinelOne, including Netskope and 1Password.

NVIDIA's technical description emphasizes continuous, in-silicon monitoring. The model is meant to give defenders visibility into what an agent is doing, which identity or authority it is using and whether policy limits remain enforceable during execution. SentinelOne said its runtime security technology will contribute behavioral visibility and controls for agent activity.

That approach addresses a basic weakness in agentic AI security. Model-level safeguards can be bypassed by prompt injection or poisoned context, while agents may hold credentials and permission to act across enterprise systems. Controls below the model can provide another enforcement layer when instructions fail.

“Today, you can't responsibly hand consequential work to an AI agent without knowing three things: what it's doing, whose authority it's acting on, and whether the limits on that authority actually hold,” SentinelOne co-founder and CEO Tomer Weingarten said.

The platform is a framework and partner ecosystem, not evidence that those guarantees have already been achieved across products. Enterprises evaluating it should ask how policy is enforced, how agent identities are traced across tools, what telemetry is retained and whether controls continue to work when a model or connector is compromised.

More technical details are available from NVIDIA Developer and SentinelOne's implementation overview.