The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md direct the AI agent to execute shell commands via the Bash tool that incorporate unvalidated user input. Specifically, in Phase 2, the placeholders <intent>, <key>, and <value> are directly embedded into node command strings used to update the workspace configuration. Since these values are sourced from the user or inferred from context without escaping or validation, an attacker could provide inputs containing shell metacharacters such as semicolons, pipes, or command substitutions (e.g., ; rm -rf /;) to achieve arbitrary code execution on the host system.

bootstrap