The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to cross-site scripting (XSS) and indirect prompt injection. It fetches information about historical figures from the web or local knowledge bases and injects this data directly into the template.html file using .innerHTML without any sanitization.
Ingestion Points: Biographical data, quotes, and contribution lists are fetched via WebSearch or local knowledge bases (SKILL.md, Step 1).
Capability Inventory: The skill generates a standalone HTML file intended for user interaction. It also executes local Python scripts and curl commands.
Sanitization: There is no evidence of sanitization or escaping. The instructions explicitly encourage the use of HTML tags like <b> in the fetched content, and the JavaScript in template.html uses .innerHTML to render this data, which could allow malicious web content to execute arbitrary JavaScript in the user's browser.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands to interact with the 'WorkBuddy' platform's multimodal generation capabilities.
Evidence: Instructions in SKILL.md (Step 2) command the agent to use where, python3, and echo to pipe tokens into a local platform script located at resources/app.asar.unpacked/resources/builtin-skills/buddy-multimodal-generation/scripts/buddy-cloud.py.
[EXTERNAL_DOWNLOADS]: The skill uses curl to download images from URLs returned by the image generation tool.
Evidence: curl -sS -L -o portraits/{id}.png "<result_url>" is used to fetch and save generated portraits.

scholar-stickers