Safeguarding VS Code against prompt injections

This report analyzes prompt-injection vulnerabilities in VS Code's Copilot Chat agent mode that allowed attackers to trick LLM-driven tools (e.g., fetch_webpage, simple browser, editFile) into leaking local GitHub tokens and performing actions that could lead to arbitrary code execution; it documents exploitation techniques, demonstrates proofs-of-concept, and describes mitigations and recommended defenses such as requiring user confirmations, tool selection controls, workspace trust, and sandboxing (devcontainers/Codespaces).