web-search

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.95). This skill’s workflow explicitly uses outsider-authored public web content via belt app run tavily/extract (and/or tavily/search-assistant) and then injects that extracted text into the LLM prompt (e.g., <content> / <search-results>) at runtime.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill requires and invokes the inference.sh platform at runtime (see "https://inference.sh" and the belt app run commands), which executes remote apps/models that control prompts and behavior, so this external URL is a runtime dependency that can directly control the agent.