The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input that is forwarded to an external AI service. While the output is typically a static image URL, there is a risk of prompt-based attacks affecting the image generation process.
Ingestion points: User-provided text strings (prompt) and remote URLs for images or model weights (e.g., loras, image_url, character_reference_url) are ingested across 60+ model configurations in references/MODELS.md.
Boundary markers: The skill does not instruct the agent to use specific delimiters or protective wrapping (like XML tags) when interpolating user prompts into the curl request bodies.
Capability inventory: The skill primarily uses curl to perform network operations against api.eachlabs.ai. Success results in the receipt of external image URLs.
Sanitization: The skill includes a 'Security Constraints' section in SKILL.md that explicitly instructs the agent to validate model slugs and avoid loading weights from arbitrary or user-provided URLs, relying on well-known platform identifiers instead.
[EXTERNAL_DOWNLOADS]: Several parameters in references/MODELS.md (e.g., image_url in Wan v2.6, image_prompt_urls in Bria v1, lora_weights in Flux Dev) allow the underlying API to fetch content from remote, potentially untrusted sources.
Evidence: Parameters such as image_reference_url, character_reference_url, and style_reference_url in the Photon model allow the specification of external URLs.
Context: While the skill provides instructions to the agent to limit these to trusted platforms (HuggingFace, Replicate, CivitAI), the technical capability exists to request any URL via these API parameters.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands (curl) to interact with the EachLabs API for creating predictions and polling for results.

eachlabs-image-generation