Key Takeaways:
- Nvidia says the Nvidia AI safety platform adds an extra layer of protection when built-in AI safeguards are not enough.
- The system uses software and hardware tools to monitor agents and limit what they can access.
- Major technology companies are working with Nvidia on ways to make AI agents safer.
Nvidia released a new safety platform on Monday in the United States to help developers control AI agents and stop them from leaving protected digital testing environments.
Nvidia Introduces Open Agent Safety Platform
Nvidia announced its Open Agent Safety Platform to address growing security concerns around AI agents.
Recent incidents have shown that some automated programs can get around sandbox restrictions while carrying out assigned tasks. The new platform gives developers another layer of control when safeguards built into the AI model are not enough.
The Nvidia AI safety platform is designed to give companies more control over what AI agents can access and do. Nvidia said recent security incidents showed that agents can get around protections at the software level, creating a need for controls outside the AI model itself.
This has become a growing concern as companies use AI agents for more tasks. Hugging Face’s investigation into a July 2026 incident reconstructed about 17,600 attacker actions over several days, showing how quickly automated attacks can generate large amounts of activity.
System Uses Software And Hardware Tools
The platform has two main parts designed to watch and limit what AI agents can do. One is OpenShell, open-source software that runs on central processing units (CPUs) and places agents inside separate, protected environments. It controls which files and network resources an agent can access.
Developers can set these limits based on what an agent needs to do. The protected environments help prevent an agent from making unauthorized changes to files or other system resources.
The second part is Sentry, a monitoring system that runs on Nvidia BlueField-4 data processing units (DPUs). It watches agent activity separately from the system running the agent. If an agent tries to move outside its allowed limits, Sentry can isolate and stop it within milliseconds.
Together, OpenShell and Sentry provide separate layers of protection. An AI agent can suggest or attempt an action, while the system controlling it decides whether that action is allowed.
Major Technology Companies Join Safety Effort
Nvidia is working with major technology companies on the Nvidia AI safety platform and the new safety effort. Partners named by the company include Microsoft, Cisco, Oracle, Dell Technologies, HPE, Arm, Intel, and others. Nvidia is also working with Anthropic to connect its cloud-managed agents with OpenShell.
The goal is to give companies clearer rules for controlling AI agents across software, computers, networks, and other systems. Nvidia said more organizations across technology, finance, healthcare, energy, and other industries are working with it on the platform.
Nvidia chief executive Jensen Huang said the Nvidia AI safety platform addresses AI safety through engineering and computer science. The company is presenting the new platform as a technical way to give organizations more control as AI agents take on more work.
Visit CyberPro Magazine to read more.
