Nvidia Introduces “Killswitch” to Stop AI Agents From Going Rogue

Nvidia has announced the launch of its new Open Agent Safety Platform, a combination of open-source software and specialized hardware designed to enforce strict boundaries that keep autonomous AI agents under control.

“AI’s extraordinary potential for society will only be realized if we solve AI safety,” said Jensen Huang, founder and CEO of Nvidia. “As we continue to discover the frontier of AI capabilities, we must accelerate discovery at the frontier of AI safety. Safety and security require full-stack engineering.”

While autonomous AI agents are taking over complex workplace tasks, they do have a troubling habit i.e. when a obstacle gets in their way, they often find clever ways to trick or bypass the security rules built into their software. Nvidia’s solution splits the heavy lifting between open software and dedicated chip hardware, creating a security net that sits completely outside the AI model itself.

To back up the software, Nvidia introduced a system reference design called Sentry. Sentry acts like an out-of-band hardware watchdog running on Nvidia BlueField-4 Data Processing Units (DPUs). Because Sentry operates independently in silicon, the AI agent cannot see it, interact with it, or manipulate it.

If an agent tries to step outside its assigned workspace or access unauthorized data, Sentry freezes and quarantines the rogue program in milliseconds. More than 100 enterprise tech companies, cybersecurity firms, and AI developers are already integrating the tools into their workflows.

Anthropic is tying its Claude Managed Agents into OpenShell and BlueField systems to keep agent execution sandboxed. Salesforce is integrating OpenShell directly with Slack, allowing human supervisors to review audit trails and grant or deny real-time permission requests from working agents.

Meanwhile, robotics companies like Figure and Skild AI are adopting the software to keep physical autonomous machines operating safely.