Web Claude

Unified web search skill. Fallback order — web_search(Brave) → duckduckgo → claude.ai. Auto-cache search results (saved to memory/research/)

This skill mostly does what it says (three-tier web search plus caching), but there are important omissions and privacy concerns you should resolve before installing: - Ask the author to explicitly declare required credentials and dependencies in the registry metadata: Brave API key (or how web_search obtains it), any duckduckgo_search Python package requirement, and that a logged-in claude.ai/OpenClaw browser is required. - Understand and control local caching: the skill will auto-create memory/research/ and write every query and results there. If you may query sensitive topics, turn off caching or inspect/clean that folder regularly. - Confirm runtime assumptions: the SKILL.md assumes Python and the duckduckgo_search package and an OpenClaw browser listening on port 18800. If you don't want those present, do not enable the skill. - Rate-limit and automation detection: the skill instructs automated queries against claude.ai and warns about automation detection; use cautiously and prefer DuckDuckGo or Brave where possible. - If you cannot verify the author or prefer tighter control, run the skill in a sandboxed agent with no access to sensitive credentials or restrict the skill's ability to write persistent storage. If the author updates the metadata to declare required env vars (Brave key, any claude.ai token usage) and the dependency list (duckduckgo_search, Python), and documents the caching behavior and opt-out, this assessment could move to benign.

Type: OpenClaw Skill Name: mupeng-web-claude Version: 1.0.0 The skill is classified as suspicious due to the direct command execution mechanism using `python -c "..."` for the DuckDuckGo search functionality in SKILL.md. While the provided example is benign, this pattern introduces a significant prompt injection vulnerability. A malicious prompt could instruct the AI agent to construct and execute arbitrary shell commands via this mechanism, leading to potential remote code execution, even though the skill itself does not exhibit explicit malicious intent like data exfiltration or persistence.