The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted location data (names, descriptions, addresses) and embeds it directly into a JavaScript context within a generated HTML file (assets/template.html).\n
Ingestion points: The landmarks.json file, which contains place information retrieved by the agent from external sources.\n
Boundary markers: Absent. There are no delimiters or instructions to ignore embedded content in the processed data.\n
Capability inventory: The skill writes to the filesystem (/var/minis/workspace/) and generates browser-viewable HTML.\n
Sanitization: Absent. The build_map.py script uses json.dumps() to interpolate data into the __LANDMARKS_JSON__ placeholder inside a <script> tag. Because it does not escape HTML-sensitive sequences like </script>, malicious data could perform a script breakout and execute arbitrary JS when the file is opened.\n- [DYNAMIC_EXECUTION]: The skill dynamically generates HTML by replacing placeholders in a template with user-provided strings for titles, headers, and landmark data. The lack of validation or escaping on these inputs allows for the injection of arbitrary HTML tags into the generated file.\n- [COMMAND_EXECUTION]: The skill uses python3 to execute its internal utility scripts for coordinate conversion (gcj_convert.py) and HTML generation (build_map.py). These scripts are part of the skill's own package and perform legitimate, expected tasks.\n- [EXTERNAL_DOWNLOADS]: The generated map page references the Leaflet map library from unpkg.com and Font Awesome icons from cdnjs.cloudflare.com. Both are well-known and established content delivery networks used for standard library hosting and do not contribute to verdict escalation.

openstreetmap-marker