The Agent Skills Directory

[SAFE]: The skill implements a comprehensive safety model in scripts/content_safety.py designed to mitigate indirect prompt injection. It uses regular expressions to detect and strip common jailbreak and instruction-override patterns in both English and Chinese from search titles and snippets.
[SAFE]: URL handling is restricted via an allowlist that only permits http:// and https:// schemes, preventing potential attacks involving javascript:, data:, or file: URIs.
[SAFE]: The skill explicitly avoids fetching full third-party page content at runtime, relying only on search snippets provided by the API engines. This significantly reduces the attack surface for malicious content ingestion.
[SAFE]: Configuration management for API keys follows standard practices, reading from environment variables or a local .env file within the skill's own directory (~/.agents/haiyuan-ai/).
[SAFE]: Persistence is limited to a local SQLite-based cache and configuration files used for intended functionality. No unauthorized persistence mechanisms (like shell profile modification) were found.
[SAFE]: Analysis of the Python dependencies (aiohttp, ddgs) and search engine client implementations (brave_client.py, tavily_client.py, exa_client.py, ddgs_client.py) shows standard usage of official APIs without suspicious behavior.

agent-search