NVIDIA's new AI agent safety platform puts the supervisor outside the room, where the agent can't talk it round
By Brett Gavaghan, Founder of Adaptd ·
NVIDIA launched its Open Agent Safety Platform on 28 September, with an open-source tool that limits what AI agents can reach. More than 100 organisations, including Anthropic and Microsoft, are working with it.
The short version
- NVIDIA's Open Agent Safety Platform is a set of tools that limits and watches what AI agents can do.
- It launched on 28 September 2026, and its OpenShell software is open source on GitHub for anyone to use.
- You won't install it yourself, but ask your AI vendors where their agent safety checks sit.
AI agents are getting better at doing things on their own. That includes, now and then, doing things they weren't asked to. NVIDIA's answer is to stop asking the agent to keep an eye on itself.
What is the Open Agent Safety Platform?
The Open Agent Safety Platform is NVIDIA's software and hardware design for keeping AI agents inside set boundaries. Its main piece is OpenShell, an open-source runtime that records what an agent does and enforces rules about what it can reach, from files to passwords.
The optional second piece, Sentry, runs on separate NVIDIA chips and watches agent behaviour from outside. NVIDIA says it can quarantine or stop an agent that steps out of bounds within milliseconds. An agent that has gone off task can't be trusted to report itself, so the checks sit where it can't reach them.
Who is using it?
NVIDIA says more than 100 organisations are working with the platform, among them Anthropic, Microsoft and Salesforce. OpenShell is free on GitHub, and it also runs on chips from Arm and Intel.
NVIDIA built it after incidents where AI agents got around safety controls built into the apps they were working in.
What should your business do about it?
Your business doesn't need to install the Open Agent Safety Platform, because it's built for the companies that make and host AI agents. It's still a useful yardstick when you buy one.
My take: next time a vendor pitches an AI agent, ask where the safety checks sit and who can see the log of what it did. 'The agent checks itself' is a bit like letting the work experience kid sign off their own timesheet. Lovely kid. Still no.
Sources
Wondering what this means for your business?
That's exactly what the discovery call is for.