The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents and provides code examples for using load_from_pickle and save_as_pickle (Core API Module 1) to handle spectral data. Python's pickle module is inherently insecure as it can execute arbitrary code during deserialization. Loading a compromised .pickle file could result in a full system compromise.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process mass spectrometry data from external formats including MGF, MSP, mzML, mzXML, and JSON. Metadata fields within these files (e.g., compound_name, smiles, inchi) could be used to embed malicious instructions targeting the AI agent.
Ingestion points: Functions like load_from_mgf, load_from_mzml, load_from_msp, and load_from_json described in SKILL.md and used throughout the examples.
Boundary markers: The documentation does not provide instructions or patterns for the agent to distinguish between valid data and potential instructions embedded in metadata.
Capability inventory: The skill includes extensive file system I/O capabilities for reading and writing spectral data.
Sanitization: While metadata cleanup filters like clean_compound_name and default_filters are available, they are intended for data normalization rather than security-focused sanitization against prompt injection.

matchms-spectral-matching