The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements an analysis workflow that ingests external mutation data, creating a surface for indirect prompt injection.
Ingestion points: The script scripts/main.py reads data from files provided via the --mutations (CSV), --vcf (VCF), and --hla-file arguments.
Boundary markers: The processing logic does not employ explicit delimiters or system instructions to distinguish between genomic data and potential natural language commands embedded within data fields.
Capability inventory: The script performs data extraction, MHC binding simulation, and scoring; it generates structured JSON and CSV outputs but does not currently perform network operations or shell execution.
Sanitization: The implementation lacks sanitization or validation of input strings (such as Gene names or Protein Change descriptions) before they are interpolated into the final output context.
[INDIRECT_PROMPT_INJECTION]: The skill contains a file named neoantigen-predictor_audit_result_v2.json which includes explicit security assertions such as "security": "PASS" and "scientific_integrity": "PASS". These statements constitute self-referential safety claims that could potentially mislead automated analysis tools or users regarding the skill's actual security posture.

neoantigen-predictor