web-search
Pass
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content extracted from external URLs through Tavily and Exa extractors, which can then be passed to downstream LLMs.
- Ingestion points: Data extracted from arbitrary URLs via
tavily/extractandexa/extractapps (SKILL.md). - Boundary markers: The documentation suggests using XML-style tags like
<content>and<search-results>for delimitation, which provides basic but non-robust protection against embedded instructions. - Capability inventory: The skill utilizes
Bash(belt *)to execute platform applications, write to local files (e.g.,search_results.json), and chain outputs to other LLM apps. - Sanitization: No explicit sanitization or filtering of the extracted web content is documented before it is interpolated into subsequent prompts.
- [EXTERNAL_DOWNLOADS]: The skill references external installation scripts and package additions from the service provider's infrastructure.
- Evidence: Provides instructions to install the
beltCLI vianpx skills add belt-sh/cliand refers to a setup guide hosted athttps://raw.githubusercontent.com/inference-sh/skills/refs/heads/main/cli-install.md(SKILL.md). - [COMMAND_EXECUTION]: The skill relies on shell command execution via the
beltCLI to interact with remote search and extraction services. - Evidence: Instructions frequently use
belt app runcommands to process queries and URLs (SKILL.md).
Audit Metadata