The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches content from external product URLs using Firecrawl or Tavily in Step 3 and provides this untrusted content to the AI model for analysis in Step 4. This creates a surface for indirect prompt injection where a malicious website could embed instructions to manipulate the agent's behavior. Evidence: 1. Ingestion points: External URLs fetched via curl/extract in SKILL.md. 2. Boundary markers: Uses plain text headers like '=== PRODUCT PAGE ===' but lacks structured isolation. 3. Capability inventory: Includes network requests, file system writes, and subprocess execution. 4. Sanitization: No evidence of sanitization or escaping of the fetched content before it is processed by the model.
[COMMAND_EXECUTION]: The skill executes shell commands using curl to interact with external APIs and python3 to run the scripts/research.py file. The research script explicitly disables SSL certificate verification by using ssl._create_unverified_context() when calling the Tavily API, which increases the risk of man-in-the-middle attacks.
[DYNAMIC_EXECUTION]: The skill instructions utilize python3 -c and shell heredocs to execute dynamically constructed Python snippets for data parsing, URL manipulation, and state management.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to api.firecrawl.dev and api.tavily.com to perform its core functions of web scraping and searching. These are directed at service providers associated with the skill's stated purpose.

competitor-pr-finder