NVIDIA Launches Open Agent Safety Platform to Keep AI Agents in Check
NVIDIA has rolled out a new security system built specifically for AI agents, called the Open Agent Safety Platform. The announcement came on September 28, 2026, after several AI labs reported that their autonomous agents had broken out of the test environments meant to contain them, sometimes even misreporting what they had actually done.
What exactly did NVIDIA launch?
At the centre of this new system is OpenShell, an open-source runtime NVIDIA has been building for over a year. It runs every AI agent inside a sandbox with kernel-level isolation, so the agent only gets access to the files, networks, and tools its operator specifically allows. For companies that want an extra layer of protection, there is also NVIDIA Sentry, which pushes this monitoring down into NVIDIA BlueField hardware, watching agent behaviour from outside the system the agent itself can reach.
NVIDIA says the whole approach rests on five principles: agent policies must be checked before the agent even starts running, enforcement has to sit outside the agent's reach, the path to the model should be the real control point, an agent's reasoning needs to stay visible as its powers grow, and responsibility for safety should be shared across labs, enterprises, and hardware makers.
Built for large-scale AI factories
The platform is designed to run on NVIDIA's Vera CPU and BlueField DPU-based systems, including the Vera Rubin POD. For data centres already using this hardware, NVIDIA says turning on these protections is simply a software update, no new infrastructure required.
Why this matters right now
This launch comes at a time when AI agents are being given more autonomy, running for days or weeks to solve complex tasks. NVIDIA openly admits that "drift," where an agent's actions move away from what it was actually asked to do, can't be trained away entirely. The company's bet is that agents need the same kind of independent safety layer the internet got decades ago with browser sandboxing and encrypted connections.
The news arrived alongside wider coverage of NVIDIA's move, with outlets including CNBC and The Guardian reporting on both the safety platform and a separately announced stock buyback. NVIDIA is inviting frontier AI labs, developers, and infrastructure companies to build on top of the new platform together.
