The Agent Skills Directory

Indirect Prompt Injection (HIGH): The skill teaches the agent to process and execute configuration from mise.toml and script files in mise-tasks/. These files are often part of the repositories an agent might be asked to analyze or manage.
Ingestion points: Project configuration files (mise.toml, mise.<env>.toml) and script files in ./mise-tasks/ or ./.mise/tasks/.
Boundary markers: Absent. The instructions do not include steps to verify the safety of these configuration files before execution.
Capability inventory: Includes arbitrary command execution (mise run, mise x), environment variable manipulation (mise set), and binary installation (mise install).
Sanitization: None. The skill assumes the configuration and task scripts are trustworthy.
Remote Code Execution (HIGH): The skill provides instructions for downloading and installing tools from various backends (npm, cargo, asdf, pipx) via mise install. If an agent is prompted to install a tool based on untrusted input, it could lead to the execution of malicious binaries or scripts.
Persistence Mechanisms (MEDIUM): The skill encourages the use of eval "$(mise activate [shell])" and suggests adding it to shell profiles (e.g., ~/.zshrc, ~/.bashrc). This behavior is used to maintain the tool's environment across sessions, which is a common persistence technique for development tools but should be monitored in an AI context.

mise