The Agent Skills Directory

Indirect Prompt Injection (LOW): The skill processes untrusted data from several CSV files and presents it to the agent without strong boundary markers. This allows potentially malicious instructions within the data to influence the agent's behavior.\n
Ingestion points: Data is read from data/charts.csv, data/colors.csv, data/stacks/jetpack-compose.csv, and data/web-interface.csv.\n
Boundary markers: The format_output function in scripts/search.py uses Markdown headers but lacks explicit 'ignore instructions' delimiters for the CSV content.\n
Capability inventory: The script possesses file-writing capabilities via the --persist flag.\n
Sanitization: No sanitization of the CSV content or the project name is visible in the provided files.\n- Data Exposure (LOW): The scripts/search.py script handles the --project-name and --output-dir arguments which are used to construct file paths. The absence of sanitization for directory traversal characters (e.g., ../) could allow an agent to be manipulated into writing files outside the intended directory, potentially impacting sensitive paths.\n- Unverifiable Logic (SAFE): The script references external local modules core.py and design_system.py which were not provided for analysis, preventing a complete audit of the search algorithm and persistence implementation.

ui-ux-pro-max