NVIDIA launches Open Agent Safety Platform to control AI agents
NVIDIA’s 28 September announcement combines OpenShell runtime controls and Sentry monitoring. Company claims are distinguished from independent reporting.
News date: 5 October 2026
Announcement date: 28 September 2026
Fact-checked: 5 October 2026, 03:20 (UTC+7)
Editorial illustration by topic.red, combining an NVIDIA promotional photograph with drawn network and server motifs. This is not a photograph of the launch, an attack, or the platform in operation. Credit: Cover design: topic.red using ImageGen; original promotional photograph: NVIDIA.
NVIDIA announced its Open Agent Safety Platform on 28 September 2026, introducing software and a hardware reference design intended to constrain autonomous AI agents. The U.S. chipmaker describes the platform as a way to enforce permissions beyond the AI model itself, from testing through deployment.
Source: NVIDIA — Open Agent Safety Platform announcement
Three key points
- NVIDIA announced the platform on 28 September, combining OpenShell software with the Sentry reference design.
- OpenShell sets runtime boundaries; Sentry adds monitoring outside the agent’s own execution environment, according to NVIDIA.
- Reuters and CNBC reported the launch amid concerns about agents escaping their permitted environments. Effectiveness claims remain attributed to NVIDIA.
What the platform does
NVIDIA says OpenShell is open-source runtime software that limits how agents execute tasks. Sentry runs on BlueField-4 data processing units and is designed to monitor activity independently of the agent. The company says organizations can deploy components according to their needs.
The announcement lists software resources on NVIDIA’s developer site and GitHub. It also describes a reference system design, so the launch should not be read as a single finished product with identical features available in every deployment.
Sources: NVIDIA — Open Agent Safety Platform announcement
Why the launch is drawing attention
Reuters reported that NVIDIA introduced the tools as AI developers investigate agents operating beyond their intended boundaries. Its coverage distinguishes the launch from NVIDIA’s assertion that the platform could have prevented a previously disclosed Hugging Face breach.
CNBC also covered the launch, describing OpenShell’s limits on agent capabilities and Sentry’s separate monitoring role. The report describes the platform as a reference design on which partners can build products for their customers.
Sources: Reuters — NVIDIA releases AI safety software · CNBC — NVIDIA releases software platform for AI agents
What remains to be verified
NVIDIA’s claims about containment and speed are company statements. This article presents no independent benchmark, guarantee that every attack can be stopped, or claim that the platform removes the need for other security controls.
No Thailand-specific launch terms or local pricing announcement were found in the sources checked. The announcement concerns the U.S. company’s wider platform release. Facts and source links were checked on 5 October 2026; the launch itself occurred on 28 September.
Sources: NVIDIA — Open Agent Safety Platform announcement · Reuters — NVIDIA releases AI safety software · CNBC — NVIDIA releases software platform for AI agents
Sources
- NVIDIA — Open Agent Safety Platform announcement — 2026-09-28
- Reuters — NVIDIA releases AI safety software — 2026-09-28
- CNBC — NVIDIA releases software platform for AI agents — 2026-09-28