The Agent Skills Directory

[COMMAND_EXECUTION]: Several scripts, including anysearch_adapter.py and zhihu_adapter.py, invoke external command-line tools using subprocess.run. The paths to these binaries and their base commands are derived from environment variables like ZHIHU_CLI or configuration files such as anysearch/runtime.conf. Although the adapters use restricted environments and avoid shell execution, this mechanism provides a surface for command execution risks if configuration files are compromised.
[DYNAMIC_EXECUTION]: The script firecrawl_adapter.py performs dynamic loading of the anysearch_adapter.py module using the importlib utility.
[INDIRECT_PROMPT_INJECTION]: The skill retrieves and processes untrusted search data from multiple third-party platforms (Zhihu, Weibo, X/Twitter, etc.). These ingestion points (found in aihot_search.py, weibo_adapter.py, etc.) present a surface for indirect prompt injection if search results contain malicious instructions. The skill includes warnings to the agent to treat these as discovery candidates and uses structured JSON output, but sanitization is limited to HTML stripping.
[EXTERNAL_DOWNLOADS]: The skill fetches data from several third-party APIs, including aihot.virxact.com, api.firecrawl.dev, and api.fxtwitter.com. These are utilized for real-time AI news and social media discovery.

yichen-unified-search