October 26, 2026: GitHub has made assisted permissions generally available in GitHub Copilot CLI, the GitHub Copilot app, and Visual Studio Code. When Copilot requests a command, filesystem-path, or URL permission, GitHub says a model evaluates eligible requests: actions classified as clearly safe can proceed, while other requests still require human approval. For QA teams using coding agents around test maintenance or CI failure triage, this is a workflow-control change worth validating before broad adoption.

What changed

Assisted permissions sit between manual approval and unconditional approval. Existing deny rules and enterprise restrictions continue to apply. GitHub documents that the feature can assess eligible requests to run commands, access paths, or fetch URLs; if the safety review is not confident, unavailable, or cannot evaluate a request, the normal permission flow applies.

For administrators, GitHub says enterprise-managed settings can block the broad Allow All mode while leaving Assisted available, or disable Assisted separately. Administrators can also choose a default permission mode for new sessions, subject to policy. Developers can select another mode only when policy permits it.

Why this matters for QA engineers

Test automation work often combines low-risk investigation—reading failure output or locating a test—with consequential actions such as changing assertions, executing scripts, or reaching a service URL. Assisted permissions may reduce interruption during routine work without granting blanket access, but GitHub explicitly cautions that model review can make mistakes and is not a substitute for sandboxing, access restrictions, or reviewing agent actions.

  • Test boundaries, not just happy paths: use a disposable repository to distinguish safe read-only commands from write, network, credential, and destructive-command requests.
  • Keep enforced controls authoritative: verify that enterprise deny rules remain blocked and that required approval rules behave as intended.
  • Capture evidence: retain the permission mode, policy version, command transcript, working-tree diff, and CI result for sample agent runs.
  • Review generated test changes: a permission decision only controls whether an action runs; it does not establish that a test, fix, or release outcome is correct.

What unattended CI runs need to prove

GitHub documents --assisted-approval for non-interactive Copilot CLI runs, including CI/CD workflows. In this mode, requests judged safe can proceed; requests that still need approval are denied, and Copilot receives the explanation so it can try a different approach. That creates a useful test case for scheduled QA automation: verify both a successful safe path and a blocked path, then confirm that the workflow fails or reports an actionable result rather than silently broadening its access.

Do not combine an assisted approach with casual broad path or URL permissions. GitHub notes that assisted approval does not change separate allow-all path or URL options, and warns that a failed assisted-permission enablement does not cancel an independently supplied broad tool-access option. Treat command arguments and standard-error warnings as release evidence in your pipeline tests.

A practical rollout checklist

  • Start with a non-production repository and a least-privilege test identity.
  • Run representative tasks: inspect a failed test, propose a narrow fix, run the selected test suite, and attempt a policy-blocked action.
  • Compare manual, assisted, and blocked outcomes for prompts, command history, file diffs, network access, and CI status.
  • Confirm that sandboxing, secret-scanning, code review, and required checks remain in force.
  • Define an owner and rollback path before making Assisted the default for a QA cohort.

The practical takeaway is not that AI agents are safe to run without oversight. It is that teams now have a documented, generally available middle setting to evaluate. For QA engineers, a controlled pilot should measure whether it removes routine friction while preserving the approval, isolation, and evidence controls that protect the delivery pipeline.

Sources