The Agent Skills Directory

[COMMAND_EXECUTION]: The skill dynamically resolves the path to an executable script (tzai-image) and executes it using bash. The resolution logic iterates through multiple directories in the user's home folder.
[COMMAND_EXECUTION]: The skill is susceptible to command injection because it performs direct interpolation of the <user subject> placeholder into a shell command string: bash "$ENGINE" mascot --prompt "<user subject>". This allows for execution of arbitrary commands if the user input contains shell metacharacters such as ;, `, or $().
[PROMPT_INJECTION]: User-provided data enters the execution context without any sanitization, escaping, or boundary markers, which can be exploited for indirect prompt injection.
[EXTERNAL_DOWNLOADS]: The skill documentation references fetching and installing the kedoupi/tzai-image-skill dependency from a remote registry via the npx tool.

tzai-mascot