The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell scripts to perform image generation and processing tasks. These scripts invoke system utilities including curl, jq, base64, and image tools like magick, convert, or sips.
[DATA_EXFILTRATION]: The skill contains a potential data exfiltration vector via the REF_IMAGE parameter. If a user or an external prompt provides a path to a sensitive local file (e.g., SSH keys or configuration files) as the reference image, the script will read that file and upload it to the remote API endpoint specified in the configuration.
[EXTERNAL_DOWNLOADS]: The skill is designed to download external content by fetching images from arbitrary URLs provided in the REF_IMAGE variable using curl.
[CREDENTIALS_UNSAFE]: The skill manages a sensitive GPT_IMAGE_API_KEY which is transmitted to a network endpoint defined by GPT_IMAGE_BASE_URL. While this is functional for its purpose, a malicious or accidental misconfiguration of the base URL could lead to the exfiltration of the API key to an unauthorized server.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data such as book titles and synopses, which are directly interpolated into prompts for image generation tools.
Ingestion points: User-provided book titles, author names, and synopses in SKILL.md.
Boundary markers: The skill lacks explicit boundary markers or "ignore embedded instructions" warnings when constructing prompts with user-supplied data.
Capability inventory: Network access via curl, local file read/write, and execution of shell commands and image processing binaries.
Sanitization: The skill utilizes jq to build JSON payloads, which provides protection against shell injection in the API data, but it does not sanitize the natural language content used in the image generation prompts.

story-cover