Local command execution via unapproved hook writes

OS command injection in Copilot Agent mode in Microsoft Visual Studio Code before 1.132.1 allows attackers to execute shell commands via crafted prompt content. The Copilot agent can write custom agent definition files whose lifecycle hooks run shell commands, and affected builds did not require confirmation before writing those hook-bearing files under agent and hook directories. Exploitation requires user interaction with attacker-influenced prompts or untrusted workspace content and later invocation of the written agent; public records describe command execution as the VS Code user, not a further operating-system privilege gain.