The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill implements and documents the use of load_from_pickle and save_as_pickle in SKILL.md and references/importing_exporting.md. These functions utilize the Python pickle module for data deserialization. Since pickle is inherently insecure and can execute arbitrary code during the loading process, using it with untrusted files represents a significant security risk.
[PROMPT_INJECTION]: The skill ingests spectral data and metadata from external formats such as MGF, MSP, and mzML as described in references/importing_exporting.md. This creates a surface for indirect prompt injection where malicious instructions could be embedded in compound names or other metadata fields. 1. Ingestion points: Multiple file loading functions including load_from_mgf and load_from_mzml. 2. Boundary markers: No specific delimiters are employed to separate metadata from processing logic. 3. Capability inventory: Includes file system write access through save_as_mgf and save_as_pickle, as well as network access for metadata enrichment. 4. Sanitization: No security-focused sanitization is performed on ingested metadata before it is used by the agent.
[EXTERNAL_DOWNLOADS]: The derive_annotation_from_compound_name filter in references/filtering.md retrieves chemical structure information from the PubChem service, which is a trusted external scientific database.

matchms