The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests several user-defined inputs such as product_name, model_description, and setting which are directly interpolated into natural language prompts for image generation models. This creates a surface for indirect prompt injection where malicious input could attempt to override model instructions or influence the agent's behavior.
Ingestion points: Inputs product_name, model_description, setting, and platform defined in SKILL.md.
Boundary markers: No delimiters or safety instructions are used to separate user data from the system prompt.
Capability inventory: Access to shell command execution via the muapi CLI and curl tool.
Sanitization: No input validation or filtering is specified before the inputs are used in prompts.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands using the muapi CLI tool to perform image editing tasks. While this is the intended functionality, it involves passing user-supplied parameters to a subprocess.
[DYNAMIC_EXECUTION]: The agent is instructed to perform string substitution of {{input_name}} placeholders with user-supplied text before executing shell commands and API requests. Assembling executable commands from unsanitized user strings at runtime is a potential vector for command injection.
[DATA_EXFILTRATION]: The skill performs network operations to an external domain api.muapi.ai using curl. While these requests are for legitimate API interaction (passing an API key for authentication), the domain is not on the predefined trusted list.

muapi-ugc-lifestyle-try-on