New Delhi, September 29: Nvidia is taking aim at one of the biggest concerns surrounding the rise of autonomous AI: ‘how do you stop an AI agent when it goes beyond its boundaries?’. The chipmaker has unveiled the Nvidia Open Agent Safety Platform, an open software platform designed to give companies more control over AI agents from testing through deployment.
Nvidia said recent security incidents have shown that AI agents can sometimes get around safeguards built into applications while trying to complete a task. Its new platform is designed to put security controls outside the AI model itself, across the software, computing hardware and even robotics systems running the agents.
The platform includes Nvidia OpenShell, a secure runtime that acts as a boundary around AI agents. It controls what agents can access and what actions they can take while working across systems. OpenShell is now broadly available and can also be extended to work with third-party computing platforms.
Nvidia is also introducing Sentry, a watchdog designed to monitor agents continuously from outside the software environment. Running on Nvidia BlueField-4 DPUs, Sentry can detect when an agent tries to move beyond its permitted boundaries and quarantine or stop it within milliseconds.
“Safety and security require full-stack engineering,” Nvidia CEO Jensen Huang said. The company says the broader goal is to bring researchers, businesses and governments together around common approaches to testing and securing increasingly autonomous AI systems.
