The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides instructions to bypass model safety filters by setting pipe.safety_checker = None. This is documented in SKILL.md, references/advanced-usage.md, and references/troubleshooting.md as a way to avoid issues like black images or for custom pipelines.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user prompts for image generation, which can be exploited via indirect prompt injection.
Ingestion points: User prompts and negative prompts in the image generation functions (e.g., pipe(prompt=...)).
Boundary markers: Absent. User input is directly passed to the diffusion pipeline without sanitization delimiters.
Capability inventory: The skill can generate images, save them to the local file system (image.save), and serve them via a FastAPI web server.
Sanitization: Absent. The documentation specifically suggests disabling the safety checker, removing the primary sanitization layer for this type of multimodal output.
[EXTERNAL_DOWNLOADS]: The skill installs standard machine learning libraries and downloads models from the HuggingFace Model Hub.
Libraries include diffusers, transformers, accelerate, torch, xformers, and peft.
Models are fetched from well-known HuggingFace repositories (e.g., stabilityai/stable-diffusion-xl-base-1.0). These are well-known services and do not elevate the verdict.

stable-diffusion-image-generation