Copilot assisted permissions: a model decides what is safe

· 6 min read

#ai

GitHub has made assisted permissions generally available in Copilot CLI, the Copilot app and Visual Studio Code. Instead of asking you about every command, a model reviews each request, lets the clearly safe ones through and asks you about the rest.

In short

Three modes, one question: who decides?

The question behind every permission prompt is who gets to say yes. The docs describe three answers:

Manual Assisted Allow-all (--yolo)
Who decides You, for each request A model for clearly safe requests, you for the rest Nobody: no review, no prompts
Deny rules Apply Apply Apply
Enterprise limits Apply Apply Apply
CLI /permissions manual /permissions assisted --allow-all or --yolo
What the docs say Shell commands can do anything your account can Not a replacement for a sandbox Only in an isolated environment

The middle column is the one that changed. Everything in the other two was already there; assisted gives you a way to stop being asked about the boring requests without giving up the question entirely.

Three Copilot permission modes: who decides Three boxes from left to right. Manual: you approve each request. Assisted: a model approves clearly safe requests and asks you about the rest. Allow-all, also called yolo: no review and no prompts. A strip underneath says deny rules and enterprise limits apply in every mode. Manual You approve each request Assisted Model approves safe, asks you the rest Allow-all aka --yolo no review, no prompts Deny rules and enterprise limits apply in every mode
The three Copilot permission modes, from most asking to least. Based on GitHub's docs on safe commands and on allowing tools.

Deny rules deserve a note: --deny-tool beats the allow options, including --allow-all. So the rules you write for the things you never want Copilot to do keep working, whichever mode you pick.

What still reaches you in assisted mode

The docs list what comes back to you: requests the model is not sure about, requests to get around the local sandbox, and cases where the review is unavailable or cannot judge the request. Requests that are already allowed run without review.

Two things the docs leave open. "Safe" has no fixed list: the decision depends on the model's confidence in each request. And the page does not say what request data goes to the review model. It does say that the review model may differ from the model you chat with, and that if it is unavailable the session's active model is used.

In autopilot and other non-interactive runs there is nobody to ask. Requests that would need your approval are denied, and the reason goes back to Copilot so it can try something else.

Turning it on

One detail for scripts and CI. --assisted-approval overrides --allow-all-tools, but it does not change --allow-all-paths or --allow-all-urls. And if assisted mode cannot be enabled, Copilot writes a warning to stderr, and --allow-all-tools can still approve a tool without any review. If you rely on this in a pipeline, read stderr.

Admins have two switches: permissions.disableBypassPermissionsMode: "disable" turns off only the unconditional mode, and permissions.disableAssistedPermissionsMode: true turns assisted off. With enterprise-managed settings they can disable Allow All and keep Assisted, and they can set the default permission mode for new sessions. If the default is not available, the session starts with manual approval.

Where it came from

The feature did not appear out of nowhere. It had a quieter start in another editor, and one other approval-related change landed in between:

Date What Where
22 Sep 2026 "Assisted approvals": low-risk tool calls approved automatically (public preview) Copilot for JetBrains 1.18.0
1 Oct 2026 Computer use: Copilot can drive desktop apps, asks before controlling one (public preview) Copilot CLI and Copilot app
26 Oct 2026 Assisted permissions (GA) Copilot CLI, Copilot app, VS Code

The JetBrains preview was the same idea under a different name: low-risk tool calls went through, high-risk ones kept asking. The computer-use preview asks before Copilot controls an app, and lets you review and reset the apps you marked "always allow". My reading, not GitHub's wording: the more an agent can reach, the more prompts there are, and assisted permissions is the answer to prompts piling up.

The good and the not-so-good

What works better

What is missing, unclear or costs something

My take, not fact: the open question is how you check a model's approvals afterwards. The pages I read document neither what "safe" means nor what the review model sees, so I would treat assisted as a convenience layer, not a security boundary.

What to do

  1. Write deny rules for the commands you never want run. They win in every mode.
  2. For risky work, keep using an isolated environment. The docs recommend it for allow-all and say assisted does not replace it.
  3. In CI, use --assisted-approval deliberately and watch stderr for the fallback warning.
  4. If you manage a team, decide whether to disable Allow All with permissions.disableBypassPermissionsMode, and whether you want Assisted on at all.

Sources