The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user data via the script and environment fields, which are interpolated directly into the prompts for the GPT model (Step 1) and the Seedance video model (Step 3). The instructions lack boundary markers or sanitization logic to prevent the user-supplied text from potentially overriding the intended prompt behavior.
Ingestion points: Input parameters {{script}}, {{environment}}, {{person}}, and {{product}} in SKILL.md.
Boundary markers: Absent in Step 1; Step 3 uses simple double quotes for {{script}}, which are easily bypassed.
Capability inventory: The skill possesses capabilities for file upload (muapi upload), image manipulation (muapi image edit), video generation (muapi video from-image), and general shell execution via curl and CLI commands.
Sanitization: No sanitization or validation of the input content is performed before interpolation.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using the muapi CLI and curl to interact with external APIs. While these are necessary for the skill's primary purpose, they establish a command execution pattern that relies on the agent's environment to handle argument security.

muapi-ugc-video-factory