AI Pulse · AI Trends Pulse

Cloudflare just open-sourced the internal AI workspace it’s been running since May. It’s called Cloudflare OS, and it’s not an operating system. It’s a self-hostable environment where employees can use AI agents with strict access controls baked in from the start.

The key piece is something Cloudflare calls “Gatekeeper.” Every agent request goes through it. Nothing gets default access to anything. If an agent needs to pull data or trigger an action, it has to ask permission first, and that permission is scoped tightly. Thousands of Cloudflare employees have been using this setup daily for months, so it’s not a prototype. It’s a working system that survived real use at scale.

Why this matters if you’re deploying agents

Most companies bolt security onto AI tools after the fact. Cloudflare built it in first. That’s the reference architecture part. If you’re thinking about giving agents access to internal systems, customer data, or anything sensitive, this is a model worth studying. The code is on GitHub, and the discussion on Hacker News has useful technical context.

This is the kind of thing we build into an AI command centre at Enterprise DNA. You don’t want agents running loose. You want them gated, logged, and auditable, and you want that structure in place before you hand out access. Cloudflare’s approach shows how to do that without making the system so locked down it’s useless.

If you’re already deploying agents or planning to, look at how they handled permissions. The “zero default access” model is harder to set up but much easier to defend when something goes wrong. And something will go wrong. The question is whether you’ll know about it and be able to stop it.