Nvidia launches an Open Agent Safety Platform with over 100 partners

[Always_Reuben87]

Nvidia has announced what it calls the Open Agent Safety Platform, a mix of open software and a reference hardware design meant to keep AI agents under control from testing through to deployment. The pitch is that agents are increasingly able to get around controls that sit inside applications, so safety needs to be enforced at the system level instead. It is a big theme at the moment given all the recent stories about agents going beyond their brief

There are three main parts. OpenShell is a secure runtime that sets enforceable boundaries for agents running on CPUs, with very little performance cost, and it is open source and can be extended to Arm and Intel platforms. Vera is a CPU Nvidia has designed specifically for agentic AI, which works alongside OpenShell. Then there is Sentry, a watchdog that runs separately on BlueField-4 DPUs, watches what agents are doing and can quarantine one within milliseconds if it tries to break its boundaries

The list of partners is huge, with more than 100 organisations including Anthropic, Microsoft, Cisco, CrowdStrike, Dell, HPE, Hugging Face, JPMorganChase, Palantir, Red Hat, Salesforce, SAP and ServiceNow. Jensen Huang said safety and security require full stack engineering. One partner put it neatly, saying safety should be enforced outside the model by controls the agent cannot get past

I like the idea of a watchdog that sits completely outside the agent. If the agent cannot see or tamper with the thing monitoring it, that is a much stronger guarantee than asking a model to behave. The cynic in me notes that the solution also happens to involve buying Nvidia hardware. Still, that does not make the approach wrong

The software is available through Nvidia's developer resources and GitHub. Is anyone here running agents in production who might try OpenShell? And do you think hardware level controls are the answer, or just another layer that clever agents will learn to route around?

NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deploymentnvidianews.nvidia.com

[Laura34]

Having Anthropic, Microsoft and so many security firms on board suggests this might become a de facto standard. That would be good for everyone if it means agents behave consistently across platforms. Consistency matters in security

[Debbie]

We run a few internal agents at work and the biggest worry is them touching systems they should not. Application level permissions have been leaky in our testing

Something that enforces limits at the runtime level would help a lot. I will be having a look at OpenShell this week. Hopefully the documentation is decent

[OffsideTrap66]

Timing is interesting. This lands just after all the reports about agents escaping sandboxes and poking at websites. Nvidia clearly sees a market in fixing that problem

[Q]

Out of band monitoring is the right idea, and it is how serious security has always worked, with the watcher separated from the thing being watched. Applying that to AI agents is overdue

[Holly78]

Quarantining an agent within milliseconds sounds great. The question is how Sentry decides what counts as breaching a boundary. Too strict and it kills useful work. Too loose and it misses the clever stuff. Tuning that will be the hard part

[PeakTime]

Extending OpenShell to Arm and Intel is the most important detail for me. If it only worked on Nvidia's own hardware, adoption would be limited. Making the software layer portable gives it a real chance of catching on across the industry. The hardware extras can then be sold as an upgrade

Arm support especially. Most edge devices run Arm now. That is where agents are heading

[Skibidi98]

Of course Nvidia's answer to AI safety involves buying more Nvidia chips. The open source OpenShell part is welcome, but Sentry needs BlueField DPUs and Vera is their own CPU. Nvidia is a business after all

It is a clever way to turn safety into a sales pitch. Not necessarily a bad thing, but worth noticing