Skip to content
Microsoft Security· Uri Oren·Indexed by this site · Originally published· Attacks →· Its place in the scan record →

Microsoft lab simulation: after device code phishing is approved, Copilot CLI sessions can be remotely controlled

The risks of device code authentication for AI coding agents

AI summary

Microsoft's security team demonstrated in a controlled experiment that device code phishing against GitHub Copilot CLI can, after user approval, lead to four abuse paths: remotely controlling a session with remote control enabled and executing commands, writing misleading memory with the /memory command, reading cloud-synced session records, and accessing private repositories accessible to that account. The attacker initiates an OAuth device flow request and tricks the user into approving it on the genuine GitHub verification page; the legitimate domain cannot indicate who initiated the request. Microsoft stresses that these are proof-of-concept results, that each impact depends on account permissions and enabled features, and that none has yet been observed in the real world; the technique is also not limited to GitHub, as any public client supporting device flow can be abused in the same way.

Read originalmicrosoft.com