Risks of device-code auth for AI coding agents
๐ This post examines how device code phishing can be used to hijack OAuth sign-ins for AI coding agents such as GitHub Copilot CLI. Microsoft simulated attacker-initiated device-code approvals in a lab to show four potential outcomes, including remote session steering, memory manipulation, synced session exposure, and repository access. The write-up describes detection and disruption capabilities in Microsoft Defender and recommends organizational controls and user vigilance to mitigate the risk.
