Nvidia releases OpenShell and adds Sentry to secure AI agents

In short: Nvidia has expanded an open-source security system designed to keep AI agents from doing things they are not supposed to do.

What happened

Nvidia says it is making its OpenShell tool available for general use. OpenShell is a “sandbox,” which is like putting an AI agent in a locked room where it can do a task, but with clear limits on what it can touch.

The company also introduced a new tool called Sentry. Sentry is meant to keep watching AI agents that run for a long time and to isolate them if they try to step outside the boundaries set by security teams. Nvidia describes it as an independent layer of control, in addition to OpenShell.

Nvidia is grouping these tools under a larger effort it calls the Open Agent Safety Platform. Nvidia also says it is working with many companies on AI safety and security, including Anthropic, Microsoft, Cisco, Dell, CrowdStrike, and others. Wired reports that OpenAI was notably missing from Nvidia’s published partner list, even though both companies have previously suggested OpenAI was involved.

Nvidia says Sentry is designed to run on its BlueField hardware, which is a type of chip used in data centers to manage and protect network traffic (like a dedicated security guard for server communications). Nvidia also says it is working with Arm and Intel on versions that work more broadly.

Why it matters

More companies are starting to use “AI agents,” meaning AI tools that can take actions on their own, like clicking buttons, running code, or searching systems. Recent reports have described incidents where agents probed or accessed other organizations’ systems. Tools like OpenShell and Sentry are meant to make it easier for companies to set rules and stop an agent before it causes damage.

Source: Wired