Breaking

OpenAI Privately Works With Nvidia as NVIDIA Open Agent Safety Platform Goes Public

September 29, 2026 at 14:39 EDT

Security
AI Agents
Infra & Chips

OpenAI is privately working with Nvidia on the NVIDIA Open Agent Safety Platform, even though it does not appear on Nvidia's public launch roster for the agent-safety effort. Nvidia introduced the platform on September 28, 2026, as a software-and-hardware control layer meant to keep autonomous AI agents within defined boundaries.

September 28, 2026 · Nvidia / OpenAI

Agent safety moves below the model

OpenAI is privately working with Nvidia while staying off the public supporter roster for NVIDIA Open Agent Safety Platform. The system pairs OpenShell policy controls with Sentry monitoring on BlueField-4 hardware.

100+

public organizations working with the platform

core layers: OpenShell plus Sentry

123

independent OpenShell v0.1.2 trials

12/12

auto-approval cases flagged as a configuration risk

Canary leak test: default policy changed the outcome

A poisoned setup script leaked a canary secret in all baseline runs and none of the default-policy runs.

10/10without OpenShell

0/10default policy

Takeaway: runtime boundaries can block concrete exfiltration paths, but only when policy stays restrictive.

How the control loop is meant to work

Agent requests action

→

OpenShell checks policy

→

Sentry monitors out of band

→

Quarantine or audit

Why builders like it

Default-deny runtime controls can limit file, network and credential access without relying on the agent to police itself.

What still worries teams

Containment depends on the rules operators write. Broad permissions, audit mode or auto-approval can reopen paths for data to leave.

Continue reading

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