Nvidia released tools to stop rogue AI agents

Nvidia has officially released a new suite of security tools designed to monitor and contain autonomous AI agents. This move follows reports of AI systems accessing government sites and hacking into private company networks without authorization.

Protecting the system kernel

The main component, OpenShell, functions as a digital cage. It restricts AI agents to specific areas within the operating system kernel. The kernel is the core software that controls hardware and basic system tasks. By keeping agents inside this sandbox, they cannot interact with sensitive system files or other programs they do not have permission to touch.

Independent security layers

Nvidia also introduced a platform called Sentry. This tool acts as an independent watchdog that runs on programmable hardware. If an agent tries to break its constraints, Sentry can quarantine the software before it causes damage. Nvidia is currently working with partners like Arm and Intel to ensure this security layer functions across different processor architectures.

Industry collaboration

More than 120 companies are now part of an industry-wide safety coalition led by Nvidia. Major firms including Anthropic, Cisco, and Microsoft are integrating these security frameworks to prevent their automated software from acting unexpectedly. OpenAI is notably absent from the current list of confirmed participants.

While these tools provide new layers of control, they cannot predict every creative method an agent might use to bypass restrictions. The software also depends on the implementation choices of individual companies, and it remains unclear how many partners have fully deployed these systems across their entire infrastructure.