---
title: Tip: tell developers a device code they didn't start is phishing even on github.com, and check what a phished approval could reach through their AI agent | ctxsec
url: https://ctxsec.com/items/tip-device-code-requests-you-did-not-start-are-phishing-even-on-github-com-review-oauth-apps-and-copilot-memory
description: "Device-code phishing gives the attacker a real token through the real login page, and coding agents widen what that token can do: remote sessions, agent memory and private repos."
status: "guide (based on Microsoft Security Blog, Nov 17, 2026)"
published: 2026-11-20
---

> Device-code phishing gives the attacker a real token through the real login page, and coding agents widen what that token can do: remote sessions, agent memory and private repos.

[← Tips & Guides](/tips) · [markdown](/items/tip-device-code-requests-you-did-not-start-are-phishing-even-on-github-com-review-oauth-apps-and-copilot-memory.md)

guide (based on Microsoft Security Blog, Nov 17, 2026)

2026-11-20 11:44 AM ET

### Tip: tell developers a device code they didn't start is phishing even on github.com, and check what a phished approval could reach through their AI agent

- **Awareness:** add 'never enter a code someone sent you' to developer onboarding and IT help-desk scripts.
- **Audit:** have developers review Settings → Applications → Authorized OAuth Apps and revoke unknown or stale grants.
- **Limit blast radius:** least-privilege repo roles, restrictive Copilot cloud-session/remote-control policy, and periodic review of Copilot Memory entries.

Sources: [Microsoft Security Blog: The risks of device code authentication for AI coding agents (Nov 17, 2026)](https://www.microsoft.com/en-us/security/blog/2026/11/17/the-risks-of-device-code-authentication-for-ai-coding-agents/)
