The Agent Skills Directory

[COMMAND_EXECUTION]: The helper script scripts/codex_imagegen_batch.sh and the instructions in SKILL.md are vulnerable to shell command injection. User-supplied prompts are interpolated directly into a double-quoted string within a bash command (codex exec ... "... '$prompt' ..."). If a prompt contains shell metacharacters such as $(...) or `...`, they will be executed by the local shell before the command is passed to the Codex CLI.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input (image prompts) and interpolates them into executable commands.
Ingestion points: The prompt variable in scripts/codex_imagegen_batch.sh and the <프롬프트> placeholder in SKILL.md.
Boundary markers: Absent. Prompts are placed directly into bash command strings.
Capability inventory: The skill executes codex exec via the shell, which has access to the workspace and the network (via the Codex tool itself).
Sanitization: Absent. There is no escaping or validation of the prompt content before it is used in a shell command.

codex-image