Breaking

NVIDIA Launches Open Agent Safety Platform to Put Runtime Limits on AI Agents

September 28, 2026 at 12:42 EDT

AI Agents
Security
Open Source

NVIDIA launched the NVIDIA Open Agent Safety Platform on September 28, 2026, positioning it as a full-stack control layer for AI agents that can keep permissions in force while agents are running. The package pairs the open-source NVIDIA OpenShell runtime with the NVIDIA Sentry reference design, moving safety enforcement outside the model and agent harness rather than relying only on prompts or model refusals.

Sep. 28, 2026 · NVIDIA

Agent safety moves outside the model

NVIDIA's Open Agent Safety Platform pairs OpenShell runtime controls with NVIDIA Sentry monitoring so agent permissions can be enforced while work is underway.

100+

organizations working with the platform technologies

80%

claimed faster sandbox performance on NVIDIA Vera

120+

organizations in the related open security alliance

Performance claim, drawn to scale

NVIDIA says Vera can deliver up to 80% faster sandbox performance than traditional CPU infrastructure.

1.0x

Traditional CPU

1.8x

NVIDIA Vera

How the boundary works

Sandbox

Agent runs inside a controlled workspace.

→

Policy check

Requests are checked outside the workload.

→

Quarantine

Sentry can isolate behavior that crosses the line.

Why teams like it

Runtime controls match how agents actually work: they need tools and credentials, but not unrestricted access to everything.

What remains hard

The strongest independent monitoring depends on BlueField-4, and policies still have to define the right limits before work starts.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.