tavily-best-practices

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 0.80). The prompt includes explicit examples that place the Tavily API key directly in a settings file and as a literal string argument (api_key="tvly-YOUR_API_KEY"), which encourages embedding secrets verbatim in code/output and therefore risks exfiltration even though an env-var option is recommended.

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

• Third-party content exposure detected (high risk: 0.95). Tavily’s required runtime workflow for extract()/crawl()/research() fetches public web pages (outsider-authored content) and ingests their readable text into the agent’s LLM context via the API response (e.g., print(response["content"]) / extracted chunks).

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

• Potentially malicious external URL detected (high risk: 0.90). The skill clearly performs runtime fetching of external pages via methods like extract()/crawl()/map() (e.g., crawl(url="https://docs.example.com")), and that fetched content is injected/used to drive LLM prompts/results, so external URLs can directly control agent behavior.