The Agent Skills Directory

Indirect Prompt Injection: The skill ingests untrusted data in the form of images and user-provided prompts (e.g., in vision-analyze.sh and the Python scripts within the examples/ directory) which are then processed by vision models. While the skill lacks technical boundary markers or input delimiters, SKILL.md provides clear operational instructions for agents to verify visual observations and avoid fabrication. The skill's capabilities include network communication with AI providers and local command execution via shell wrappers. Although image_utils.py performs image sanitization through resizing and re-encoding to manage payload sizes, these ingestion points represent an inherent attack surface typical for multimodal AI tools.
Command Execution: The skill uses shell scripts to wrap Python execution, facilitating environment setup and model interaction. This setup uses standard command-line patterns to pass arguments like file paths and prompts, ensuring the skill remains within its intended operational scope while providing a convenient interface for the agent.
External Downloads: During installation, the skill fetches widely-used SDKs (anthropic, openai, google-genai) and the pillow library from official package registries. These dependencies are maintained by established technology organizations and are necessary for the skill's primary functionality of interacting with vision models and processing images.

image-vision