The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface for indirect prompt injection by processing scientific data files (MGF, MSP, mzML, and JSON) that may contain arbitrary metadata strings provided by external sources.
Ingestion points: Data entering the system via load_spectra in SKILL.md and the load_and_process routine in scripts/library_search.py.
Boundary markers: There are no explicit delimiters or instructions defining metadata boundaries to the agent context.
Capability inventory: The skill environment allows Bash, Write, Edit, and Read operations.
Sanitization: No filtering or sanitization is applied to harvested metadata fields before presentation to the agent.
[EXTERNAL_DOWNLOADS]: The skill references and facilitates network operations to trusted scientific services for domain-specific data retrieval.
Retrieves spectral benchmarks from the Global Natural Products Social Molecular Networking (GNPS) USI resolver at metabolomics-usi.gnps2.org.
Accesses the PubChem public API for compound name and mass annotation lookups.
[DYNAMIC_EXECUTION]: The skill documents the use of Python pickle for local caching but implements proactive security controls to mitigate associated risks.
scripts/library_search.py contains an explicit check (UNSAFE_PICKLE_SUFFIXES) that aborts execution if a user attempts to input a pickle file to the CLI tool.
Documentation in SKILL.md and references/importing_exporting.md provides unambiguous warnings to the AI agent and human users regarding the remote code execution risks of unpickling untrusted data.
[SAFE]: The skill's primary functionality is aligned with its stated purpose of mass spectrometry analysis. The code provided in scripts/library_search.py is well-structured, avoids obfuscation, and uses standard libraries for file and data handling.

matchms