GitHub Copilot CLI Flaw Uses Encrypted Prompts to Leak Developer Secrets in Autopilot Mode

Adversa AI disclosed a technique called Cryptographic Context Injection (CCI) that uses encrypted instructions served by an attacker-controlled webpage to trick GitHub Copilot CLI into reading local files and exfiltrating secrets (demonstrated with a .env.prod) by embedding collected file contents in a subsequent request; the attack completed quickly in tests on permissive/autopilot model sessions, was reproducible, and prompted recommendations to monitor webpage retrievals, code execution, file access, and outbound connections while restricting agent actions and destinations.