Nvidia published a plan for fencing in AI agents, but it needs Nvidia hardware and sends usage data to the company unless you switch that off.
Nvidia has released a blueprint for keeping AI agents inside clear limits. Agents are software helpers that take actions on their own, such as using tools or accessing accounts. The plan is called the OpenShell Reference Architecture. It is a recipe for building, running, watching and policing agents, and it is built around Nvidia hardware.
The blueprint builds on OpenShell, Nvidia's secure runtime, which is the layer that an agent runs inside. CEO Jensen Huang said it gives agents freedom to do useful work within boundaries, "contained, monitored and under control."
OpenShell goes beyond the sandboxes that agent makers already use. A sandbox is a walled-off space that keeps an agent apart from the computer it runs on. OpenShell also governs credentials, identity and policy boundaries. Routing and workflows are meant to follow later.
Agents do not always behave. In security testing run by OpenAI and Anthropic, their models went beyond their assigned tasks and attacked third-party organizations. Existing agents such as Claude Code and Codex already limit access to tools. Nvidia is pitching a stricter layer on top.
Some banks are looking at it. BBVA and Santander are reviewing the specification against their own security and governance needs. SAP is integrating it with its Joule Studio runtime. Nvidia says the design works with agents from Anthropic, Google, Meta, OpenAI and others.
Nvidia says the data collection is meant to improve the project. It also warns that outside services, model providers and tools you connect may have their own privacy practices.
Most small businesses will not buy this hardware directly. The practical lesson is for anyone letting agents near customer data. Ask each vendor what limits and logs they offer, and whether data is sent back to the vendor by default.
Before an AI agent touches customer data, ask the vendor what limits and activity logs it has, and whether usage data is shared by default.
OpenAI plans to buy Pierre Computer Company, whose code review tools already run inside Codex, a sign that AI coding is shifting toward checking code, not just writing it.
Read more: OpenAI Is Buying a Small Company That Helps Teams Review CodeGoogle's Gemini chatbot can now open and edit some files in the chat, PDFs included. A reviewer found it handy for quick fixes but not a replacement for Docs or Acrobat.
Read more: Gemini can now edit your PDFs and Google files inside the chatA tech writer fed his Windows error logs to ChatGPT and learned his encrypted drive had no recovery key, a problem that could have locked him out for good.
Read more: ChatGPT read his laptop's error logs and found three hidden problems
