The Agent Skills Directory

[METADATA_POISONING]: The skill includes a file named meta-funnel-plot_audit_result_v2.json which impersonates an official security audit report.
Evidence: The file contains fields such as security: PASS, deployable: true, and grade: Production Ready. This artifact is designed to mislead security systems or human reviewers into believing the skill has undergone and passed an external security verification. This is a deceptive practice intended to bypass security scrutiny.
[COMMAND_EXECUTION]: The execution flow involves spawning subprocesses to run R and Python analysis scripts.
Evidence: In scripts/run_funnel_plot.py, the run_r_script and run_python_script functions use subprocess.run() to execute Rscript and python commands. While these commands are invoked using list-based arguments to mitigate shell injection, the reliance on runtime execution of scripts using user-provided parameters represents a significant capability surface.
[INDIRECT_PROMPT_INJECTION]: The skill contains self-referential metadata that targets the analysis process and processes external data without sanitization.
Ingestion points: meta-funnel-plot_audit_result_v2.json (self-referential) and user-provided CSV files.
Boundary markers: None. The deceptive JSON is presented alongside valid skill files, and no explicit delimiters are defined for CSV data processing.
Capability inventory: File writing (write.csv in R), file reading (read.csv in R and pd.read_csv in Python), and subprocess execution (subprocess.run in Python).
Sanitization: The skill does not sanitize strings read from user-provided CSV files before they are included in output summaries, creating a potential surface for prompt injection if results are presented to an agent.

meta-funnel-plot