What happened
Microsoft Security lab research (Nov 17) shows device code phishing, a technique that applies to any device-flow client, can be extended to AI coding agents. Using GitHub Copilot CLI as an illustrative example, a phished device-code approval let an attacker remotely steer an active agent session (run commands subject to local approval), add misleading entries to agent memory, read cloud-synchronized sessions, and access private repositories. The post details how Microsoft Defender detects these attempts by linking phishing to GitHub auth activity and how to disrupt suspicious access.
Why it matters
It identifies a practical account-takeover path into agentic CI/CD tooling and gives detection/remediation guidance, expanding awareness that standard OAuth device-flow consent is now a vector against AI agents — relevant to any organization running coding agents.
Applicability
Security and developer-platform teams running GitHub Copilot CLI / other device-flow agents; anyone deploying coding assistants and reviewing agent permissions and remote-control settings.