Payments are experiencing issues due to temporary restrictions in Russia. If your payment does not go through, please submit a support request.Our support team is available 24/7 — we are always here to help with hosting and server issues.We are now accepting requests for dedicated server rental and colocation services in our data center.Reminder: we recommend enabling backups for additional data protection.A new VPS/VDS lineup with NVMe storage and improved performance is now available.Maintenance work on some servers has been completed. All services are operating normally.
Article3 min read

AI Agent Access to a Single App: Why Cloudflare Split Permissions in Workers

New Workers roles allow restricting access to specific applications. I explain with clear examples why viewing, publishing, and deletion rights must be separated.

One key opens a single small compartment among locked modules
In this article

A developer or AI agent has a task to fix a single service. To do this, they are granted access to the entire cloud account simply because it is faster to set up. The cost of this convenience becomes apparent when an error occurs: the scope of potential damage is far wider than the assigned task.

On September 15, 2026, Cloudflare introduced more granular permissions for Workers. Access can be restricted to a specific application with a chosen role: viewing metadata, reading content, editing, or administering. These restrictions also apply to tokens for automated systems.

Viewing must not silently turn into changing

A specialist looking for the cause of a failure may only need logs and metrics. Someone studying the code requires a different level of access. Publishing a new version is a separate operation with consequences for users.

New Cloudflare roles make this distinction explicit. For example, an editor can update content and settings but cannot create or delete resources. This approach is worth adopting as a baseline for permission audits in other company systems, even if role names differ.

App-level restrictions do not replace common sense

A small service may have access to a shared customer database. In that case, granting permission to modify its code remains sensitive. You must evaluate not only the visible application but also the data and actions it can access directly.

Therefore, before granting permissions, draw a simple chain of responsibility: what task does the participant perform, what must they see, what changes are allowed, and who approves the release. This helps select minimally sufficient privileges without constant manual exceptions.

Tokens should not outlive their tasks

Automated systems need their own access, revocable independently of any person. A shared secret for the entire team complicates investigations: log entries alone often reveal nothing about who actually performed an action.

It is useful to track token assignments and integration owners. If an experiment ends, its access should not remain active for months out of habit. The process for revoking access is just as important as the initial issuance.

How to verify that boundaries work

Both allowed and denied actions are verified. The participant must be able to complete their task without gaining access to adjacent projects. Such checks are performed on safe resources, without attempts to delete the live application.

There must also be a response plan for errors: who will stop the automation, how to revoke its token, and where to review the change history. Limited rights reduce risk but do not make an incorrect update harmless.

Questions to ask your team

Start with one question: what access will the contractor receive tomorrow to fix a form? If the answer is the main administrator password, you have already found a valuable security improvement.

The Workers update signals the direction of platform evolution: permissions are becoming more task-specific. For businesses, this offers the ability to delegate work to people and automation with clear boundaries of responsibility.

Discussion0

Share your experience and ask questions. Comments without links appear after editorial review.

No comments yet. Start the discussion.