tavily-dynamic-search
Warn
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill instructs the agent to generate and execute Python code at runtime using heredocs (python3 << 'PYEOF') or command-line execution flags (python3 -c). This pattern facilitates the execution of logic dynamically created by the agent to process data.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content retrieved from the web via the Tavily API, which is ingested into a local environment for filtering.
- Ingestion points: Raw JSON and markdown content from tvly search and extract results as described in SKILL.md.
- Boundary markers: Absent; the process relies entirely on the agent's internal filtering logic to isolate relevant content.
- Capability inventory: Access to shell command execution (subprocess), local filesystem writes (/tmp/), and package installation (pip/uv).
- Sanitization: Absent; raw web data is processed directly by agent-generated scripts without prior validation.
- [EXTERNAL_DOWNLOADS]: The skill guides the installation of the official tavily-cli package from standard public registries using pip or uv.
- [COMMAND_EXECUTION]: The skill utilizes subprocess.check_output and shell piping to execute local CLI tools and orchestrate the search workflow.
Audit Metadata