Wain AI/Tech Blog

AI news and trends worldwide, updated nearly every day

GitHub Copilot's "Assisted Permissions" Reaches General Availability - A Model Decides Which Actions to Approve, and GitHub Says Enterprise Ask Rules Can Be Satisfied in the CLI

GitHub Copilot's "Assisted Permissions" Reaches General Availability - A Model Decides Which Actions to Approve, and GitHub Says Enterprise Ask Rules Can Be Satisfied in the CLI

On October 26, 2026, GitHub made assisted permissions generally available in Copilot CLI, the Copilot app and VS Code. A model reviews permission requests and lets through the ones it judges safe. GitHub's documentation says that in Copilot CLI, enterprise ask rules can also be satisfied by the model's approval.

GitHub announced in its Changelog on October 26, 2026 that “assisted permissions” has reached general availability in GitHub Copilot CLI, the GitHub Copilot app and Visual Studio Code 1. Whenever Copilot needs approval to execute a command, access a path on disk or visit a URL, an LLM evaluates that request first. Requests it classifies as clearly safe go ahead on their own; everything else still waits for the user.

GitHub’s documentation positions it between having a person approve each action and the option that allows everything 2. Administrators also get settings to keep the allow-everything mode switched off while leaving assisted permissions available, and to decide which mode new sessions open in. At the same time, GitHub’s documentation states that in Copilot CLI, the ask rules an enterprise sets can be satisfied by the model’s approval. Administrators who rely on ask rules to keep a person in the loop will want to revisit that assumption.

Three ways to approve

According to GitHub’s documentation, users can choose among three approaches 2.

ModeBehavior
Manual approvalCopilot asks before any action that is not already allowed, such as a command that is not read-only, a file edit or a URL fetch. Approval can be given one time only or for the whole session
Assisted permissionsA model assesses the requests that are eligible for review and approves the ones it judges safe. Only the requests still needing approval reach the user
Unconditional approvalTools, paths and URLs are available broadly, with neither a model review nor per-action prompts. In Copilot CLI this corresponds to --allow-all (--yolo) or /permissions allow-all (/yolo)

Turning on assisted permissions does not skip existing rules or human approvals; those are checked before the model gets involved. Anything already permitted runs with no prompt and no model review, and enterprise restrictions and deny rules remain in force. When the model is confident a request is safe, the work continues; when it is not, the user gets a prompt just as under manual approval. If no review can be performed, or the review cannot judge the request, the session falls back to its normal permission handling.

Copilot picks the model that performs the review, and it may not be the model chosen for the conversation. For users who bring their own model provider, the session’s active model does the review too. The documentation cautions that assisted permissions cuts down on prompts but does not take away the user’s responsibility for what Copilot does: the model can get it wrong and approve something dangerous or unintended. It also says plainly that the feature does not replace sandboxing or access restrictions.

To turn it on, Copilot CLI users pick assisted from /permissions, Copilot app users open “Tool permissions” in the mode menu and pick Assisted, and VS Code users select Assisted permissions in the permissions picker 1. The Copilot app also keeps that choice as the starting mode for its new local sessions, and cloud sessions do not offer tool permission controls at all 2.

Unattended runs such as CI

The most direct operational impact is on runs where nobody is around to answer prompts, such as CI/CD workflows and scheduled jobs. The documentation explains that with no permission options, anything requiring approval is refused and the task frequently cannot be completed, while --allow-all-tools hands out broad tool permissions with no model review 2.

For these runs, the in-between choice is the --assisted-approval option, or setting the environment variable COPILOT_ASSISTED_APPROVAL=true. Requests the review considers safe are approved automatically. Requests the review wants a human to approve are refused, and the reasoning is returned to Copilot, letting it attempt another route. The option only works for non-interactive runs; passing it when launching an interactive session makes the CLI exit with an error.

The documentation flags two pitfalls. First, --assisted-approval overrides --allow-all-tools, but it leaves --allow-all-paths and --allow-all-urls untouched. Combined with --allow-all (--yolo), tool requests go through the review while path and URL access may remain unrestricted. Second, if enterprise settings prevent assisted permissions from being enabled, the CLI only prints a warning to standard error and does not revoke a separately supplied --allow-all-tools. Tool requests could then be approved with no review, so the documentation asks users to watch for that warning in automated output 2. Simply adding --assisted-approval to an existing workflow that still carries --allow-all-style flags may therefore not produce the restriction the team intended.

Three things administrators can set separately

Enterprise administrators can configure three separate items in enterprise managed settings 3.

  • Setting permissions.disableBypassPermissionsMode to "disable" turns off the allow-everything mode. In Copilot CLI, --yolo, --allow-all and the individual --allow-all-tools, --allow-all-paths and --allow-all-urls flags are ignored when the CLI starts, and /permissions allow-all is blocked as well. It does not block /permissions assisted
  • Setting permissions.disableAssistedPermissionsMode to true prevents assisted permissions from being turned on by users. It is separate from the setting above and does not, on its own, turn off the allow-everything mode
  • permissions.defaultMode sets which mode new sessions begin in, from "manual", "assisted" and "allow-all". It is a default rather than a lock, so users can pick a different mode per session within what policy permits. A session that cannot use the chosen mode, or whose policy forbids it, opens in manual approval instead

The Changelog likewise describes keeping the allow-everything mode blocked while leaving assisted permissions available, or disabling assisted permissions on its own 1. In Copilot CLI, the administrator’s defaultMode wins over the user’s own default (defaultPermissionMode in ~/.copilot/settings.json). If neither has set a default, new sessions may start in either manual or assisted mode, and users are pointed to /permissions show to check 2.

The two documentation pages describe ask rules differently

What administrators need to check is how this interacts with ask rules, which require a person to approve an operation each time.

The assisted permissions page says that for Copilot CLI, the ask rules an enterprise defines for shell commands, reading or writing files, and URLs may be met by a model approval, provided the request qualifies for review. In other words, those rules will not necessarily produce a prompt for the user. Deny rules set by the enterprise are unaffected. Remaining tool calls that need approval under enterprise rules, and requests to bypass the local sandbox, are not reviewed and follow the usual permission handling 2.

The enterprise managed settings page, by contrast, describes a managed ask rule as one that neither the allow-everything mode, an auto-approval setting, hooks and similar shortcuts, nor a previously saved approval can satisfy, so a fresh prompt appears every time the operation is requested 3. That page does not say whether assisted permissions counts as such an “auto-approval setting.” Its section on disableBypassPermissionsMode also says managed deny and ask rules stay in effect. Which description reflects the actual behavior cannot be settled from the published material alone.

For organizations that designed their ask rules on the premise that a person always confirms, this is a reason to consider whether the operations they want checked should be blocked with deny instead, or whether assisted permissions should be turned off with disableAssistedPermissionsMode. At a minimum, it seems prudent to confirm in Copilot CLI whether a prompt actually appears before deciding on the default mode.

Handing approval decisions to a model

GitHub’s documentation notes that answering approval prompts keeps the user in control but can slow things down, especially early in a session while Copilot is still working out which actions it may take 2. Allowing everything to avoid prompts means no one oversees each action. Assisted permissions places a reviewing model between those two positions. A comparable mechanism elsewhere is Claude Code, where Anthropic announced on August 7 that auto mode, in which a classifier screens tool calls, would become the default for new sessions on the Pro, Max and Team plans.

As GitHub itself says the feature is no substitute for a sandbox, the model’s review is one layer of defense among several. In August, Copilot also gained a setting to narrow the MCP servers it can run with allow and deny lists. Whether to make assisted permissions the default is a decision best made together with such restrictions, after working out where things would stop if the model’s review got it wrong.

Sources

  1. GitHub Copilot assisted permissions generally available - GitHub Changelog (October 26, 2026)
  2. Allowing safe commands to run without manual approval - GitHub Docs
  3. Enterprise managed settings - GitHub Docs (permissions section)

We publish the latest AI news nearly every day.

Subscribe via RSS Get new posts the moment they go live.

Search other keywords →