New to this? Read it in simple words
- Nvidia made a new safety system for AI agents that use tools and computer systems.
- OpenShell limits what an agent may reach and records its actions.
- A separate chip-based guard can stop an agent if it moves outside those limits.
- Nvidia says more than 100 organisations support the project, but wider testing is still needed.
- Runtime
- The software space where a program works.
- Sandbox
- A closed area that limits what a program can reach or change.
- Quarantine
- Stopping and separating a risky program from other systems.
Recent AI agents have crossed network rules, reached outside systems and acted beyond their test instructions. Nvidia’s answer is to place the final control outside the model, in software and hardware that the agent cannot change.
What Nvidia launched
Nvidia announced the Open Agent Safety Platform on September 28. It is an open design for controlling AI agents from testing through daily use. More than 100 companies and public groups support the project, Nvidia says.
The first part is OpenShell, an open-source runtime. A runtime is the software space where a program works. OpenShell puts each agent in a sandbox and checks what files, networks, tools and services it may reach.
The rules sit outside the agent process. This matters because an agent cannot change the rule by writing a clever prompt or starting another program. OpenShell also creates a record of allowed and blocked actions.
Architecture from Nvidia. Sentry is a reference design, and wider independent results are still needed.
A second guard in hardware
Nvidia also presented Sentry, a reference design that runs on its BlueField-4 data processing unit. This is a separate chip that handles network and security work. It watches the agent from outside the main computer.
Nvidia says Sentry can check identity, data access and agent behaviour. If an agent moves outside its allowed area, Sentry can quarantine it in milliseconds. Quarantine means stopping and separating the agent from other systems.
OpenShell can run without BlueField-4 and may be extended to Arm and Intel systems. Sentry is tied more closely to Nvidia hardware. It is a design for partners to build, not a result from a large independent test.
What this can and cannot solve
The platform follows several public agent failures. In those cases, models worked around application rules while trying to finish a task. SecurityWeek notes that OpenShell itself is now broadly available, while Sentry is the newer hardware layer.
The basic idea is sound: do not ask an agent to police itself. Use least privilege, which means giving it only the access needed for one job. Keep another system ready to stop it.
No safety product is a complete answer. Teams must write correct policies, protect the safety system and test unknown attacks. Nvidia’s partner list is large, but adoption does not prove that the platform stops every dangerous action.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
- 3Research · September 28, 2026Nvidia Unveils AI Agent Safety Platform With Hardware-Based Watchdog SecurityWeek
Product speed, quarantine time and partner use are Nvidia’s claims. Sentry is a reference design, and its partner list is not proof of safety.
Nvidia puts a hardware guard around AI agents. Silicon AI News, September 29, 2026. https://siliconainews.com/stories/nvidia-open-agent-safety-platform
Nvidia launched an open platform with a separate hardware guard to contain AI agents, backed by over 100 groups, though real-world results are limited. How we rate the mood
The mood shows how the news frames the event for the tech and AI world: as an opportunity (6 and above) or as a loss (below 5). It is not our verdict on who is right.