The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes spectral data from various external sources, establishing a potential vector for indirect prompt injection via malformed metadata.
Ingestion points: Data is loaded from local spectral files (MGF, MSP, mzML, mzXML, JSON) and remote accessions through the Metabolomics USI service (https://metabolomics-usi.gnps2.org) as documented in SKILL.md and references/importing_exporting.md.
Boundary markers: The processing logic does not utilize explicit delimiters or specialized instructions to isolate spectral metadata fields from the agent's instructional context.
Capability inventory: The skill utilizes Bash for runtime verification and environment management, alongside Read, Write, and Edit tools for manipulating spectral libraries.
Sanitization: The scripts/library_search.py script implements robust defensive measures by strictly blocking the ingestion of .pickle and .pkl files to prevent remote code execution. It also validates command-line arguments and enforces maximum pair limits to mitigate resource exhaustion attacks.

matchms