The Agent Skills Directory

[DATA_EXFILTRATION]: The generate_poster.py script reads the content of a file provided via the --image argument and transmits it to https://giggle.pro/api/v1/generation/image-to-image via a POST request. The skill instructions in SKILL.md explicitly ask the user for a local file path, and the script lacks validation to ensure the file is an image or within a safe directory. An attacker could provide paths to sensitive files (e.g., ~/.ssh/id_rsa, .env, or AWS credentials) for exfiltration.
[COMMAND_EXECUTION]: The workflow in SKILL.md instructs the agent to execute a shell command where user-controlled strings (<product_image_path>, <composed_prompt>) are interpolated into the command line for generate_poster.py. This pattern is susceptible to command injection if the agent does not strictly sanitize or escape the user input before execution.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input and interpolates it directly into the prompt template used for image generation.
Ingestion points: User-supplied product name, selling points, and style preferences in SKILL.md Stage 1.
Boundary markers: Absent. The prompt is constructed by simple string interpolation without delimiters or instructions to ignore embedded commands.
Capability inventory: The generate_poster.py script provides file system read access and network communication capabilities.
Sanitization: No input validation, filtering, or escaping is performed on the user-provided strings or file paths.

giggle-product-poster