The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions on how to bypass safety filters that are typically enabled in diffusion pipelines to prevent the generation of prohibited or sensitive content.
Evidence: In SKILL.md, references/advanced-usage.md, and references/troubleshooting.md, the instructions recommend setting pipe.safety_checker = None and requires_safety_checker=False as solutions for technical issues like "black images" or as part of custom pipeline configurations.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest natural language prompts to generate images, creating a surface for indirect prompt injection if the inputs are sourced from untrusted external data.
Ingestion points: The skill accepts prompt and negative_prompt as string inputs across various scripts and API examples (e.g., SKILL.md, references/advanced-usage.md).
Boundary markers: No delimiters or boundary markers are suggested for the text prompts.
Capability inventory: The skill has the capability to write files to the local system (image.save) and perform network operations to download models.
Sanitization: There is no evidence of input validation or sanitization for the prompts, which could be exploited to influence the behavior of the diffusion model or potentially the agent processing the response.
[EXTERNAL_DOWNLOADS]: The skill performs downloads of large model weights and configurations from well-known and community-standard repositories.
Evidence: The skill uses from_pretrained() to fetch models from Hugging Face (huggingface.co), including stable-diffusion-v1-5, stabilityai/stable-diffusion-xl-base-1.0, and runwayml/stable-diffusion-inpainting.
Evidence: references/troubleshooting.md mentions using hf-mirror.com as a third-party mirror for Hugging Face content and installing xformers directly from a GitHub repository (github.com/facebookresearch/xformers).

stable-diffusion-image-generation