The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external files (such as SVG, PDF, and various raster image formats) provided by the user, which presents a surface area for indirect prompt injection or malicious data ingestion.
Ingestion points: The scripts/image_metadata.py script reads and parses user-supplied files to extract metadata.
Boundary markers: Not applicable, as the skill performs technical analysis of file structures rather than interpreting natural language content from within these files.
Capability inventory: The skill is configured with file read/write and bash execution permissions.
Sanitization: The implementation includes excellent defensive measures, such as rejecting SVG files containing internal DTD or entity declarations to prevent XXE attacks, and enforcing strict pixel limits using Pillow to mitigate decompression bomb (Zip Bomb) attacks.
[DYNAMIC_EXECUTION]: The skill uses dynamic loading to manage its internal assets and styles.
Evidence: scripts/style_presets.py and scripts/palette_audit.py use importlib.util to load the local assets/color_palettes.py module at runtime.
Context: This is a standard and secure practice for modularity within a local filesystem context and does not involve executing code from remote or untrusted sources.
[COMMAND_EXECUTION]: The skill scripts utilize argparse and standard library functions for local file manipulation.
Evidence: The CLI tools provided (image_metadata.py, figure_export.py, etc.) are designed to be run locally via uv.
Safety: The scripts include a dedicated _common.py module that enforces safety checks, such as rejecting symlinks for both input and output paths and performing atomic writes to prevent file corruption or race conditions.

scientific-visualization