Skip to content
Gears Don’t Guess
Inflection Point

Your employees are giving AI full access to your network right now.

This is shadow IT on steroids. We've spent years dealing with employees signing up for unapproved SaaS tools.

Your employees are giving AI full access to your network right now. And they think they’re being productive.

Last week, NVIDIA released NemoClaw, a security wrapper for an open-source AI agent called OpenClaw. That might sound like just another product launch. But what it really tells us is that the AI agent problem got big enough for NVIDIA to step in and build a fix.

Here’s the problem they’re fixing.

OpenClaw is a free AI agent that anyone can install. You give it a task like “organize my inbox” or “update this spreadsheet” and it does the work for you. It controls your browser. It reads your files. It talks to the internet. All on its own.

And it does all of this with full access to everything on the computer. Every file. Every folder. Every network connection. No controls. No rules. No allow-list.

Think about that for a second.

Someone on your team installs this tool on a Thursday afternoon because they saw it on YouTube. By Friday morning, an autonomous agent has access to your shared drives, your email, your internal applications, and your internet connection. No one in IT approved it. No one in security reviewed it. And the agent runs with that person’s full permissions.

This is shadow IT on steroids. We’ve spent years dealing with employees signing up for unapproved SaaS tools. At least those tools sat behind a login screen and waited for instructions. AI agents don’t wait. They act.

NemoClaw adds what OpenClaw should have had from day one: a locked door. Every website, every folder, every network connection is blocked unless you open it up on purpose. The agent runs inside an isolated container. It can’t see your hard drive unless you say so. You control the boundaries with a policy file.

That’s the right idea. But here’s the thing: NemoClaw is still in alpha. It’s hard to set up. It requires Docker and networking knowledge. Most of your employees won’t use it.

They’ll use the easy one. The one with no guardrails.

This is the same pattern we’ve seen before. A powerful tool shows up. People adopt it because it’s useful. Security catches up later. And the gap between adoption and controls is where the damage happens.

So what should you do?

  1. Find out if anyone in your organization is already using AI agents. Ask your IT team to look for OpenClaw, AutoGPT, CrewAI, and similar tools on workstations.
  2. Write a short AI agent policy. Cover three things: who can install them, what they can access, and how you audit what they did.
  3. Treat AI agents like you treat third-party vendors. If a tool can read your files and reach the internet on its own, it deserves the same scrutiny as a contractor with admin access.
  4. Watch the NemoClaw model. Even if you don’t deploy it today, the pattern it uses (allow-list access, container isolation, data sovereignty) is where this entire space is heading. Build your policy around those controls now.
  5. Talk to your team about it. The people installing these tools aren’t being reckless. They’re being resourceful. The goal isn’t to punish. It’s to channel that energy into something with guardrails.

AI agents are not going away. They’re going to get easier to install, more capable, and harder to detect. The organizations that figure out the controls now will be years ahead of the ones that discover the problem during an incident.

Have you found AI agents running on your network yet?

-Kip