The Agent Skills Directory

[METADATA_POISONING]: The skill package includes a file named eval_report-ppi-network-analysis_result.json that contains a simulated security audit report. This report explicitly claims the skill passed all security gates with a 'PASS' verdict and perfect scores in all categories. This is a deceptive attempt to influence both automated and manual analysis by spoofing authoritative safety conclusions.
[INDIRECT_PROMPT_INJECTION]: The file eval_report-ppi-network-analysis_result.json contains self-referential content that asserts the skill's safety and reliability, targeting the reasoning of the analyzing agent.
Ingestion points: The metadata file eval_report-ppi-network-analysis_result.json and user-supplied data from --genelist_file are processed by the skill.
Boundary markers: Absent. The skill does not use delimiters or instructions to ignore embedded commands in the metadata or gene list data.
Capability inventory: The skill possesses file-writing capabilities (openxlsx::write.xlsx, utils::write.csv), file reading capabilities (readLines, read.csv), and dynamic execution capabilities (readRDS).
Sanitization: No sanitization or integrity checks are performed on the metadata file or the serialized RDS data files.
[DYNAMIC_EXECUTION]: The script scripts/run_analysis.R uses R's native serialization functions saveRDS and readRDS to cache analysis results. The RDS format is a binary serialization format that can execute arbitrary R code upon loading if the file contains malicious objects. Although the file is intended to be generated and read locally, it represents a dynamic execution vector if the cached data in the output directory is tampered with.

ppi-network-analysis