The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/generate-nano-art.sh is vulnerable to command injection. User-supplied arguments such as --subject, --action, and others are interpolated into the EXPERT_PROMPT variable using double quotes. In bash, double-quoted strings allow for command substitution ($(...) or backticks), meaning that any command embedded in the user input will be executed by the shell during variable assignment.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to take arbitrary user input and transform it into reasoning briefs. It lacks sanitization, escaping, or boundary markers to ensure that malicious inputs cannot break out of the intended prompt structure or influence the execution environment. Evidence: 1. Ingestion points: User-provided subject, action, and context descriptions in SKILL.md; 2. Boundary markers: Absent; 3. Capability inventory: The skill executes shell scripts and calls a core media primitive (scripts/generate-nano-art.sh, core/media/generate-image.sh); 4. Sanitization: Absent.
[DYNAMIC_EXECUTION]: The skill dynamically constructs a prompt and a command-line call to a core primitive using unvalidated runtime data. This pattern facilitates the command injection vulnerability described above.

muapi-nano-banana