web-search

Warn

Audited by Gen Agent Trust Hub on Nov 18, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill requires the installation of the belt CLI tool (npx skills add belt-sh/cli) and references several other external skills from the inference-sh organization. It also links to external installation scripts and documentation hosted on a third-party GitHub repository (https://raw.githubusercontent.com/inference-sh/skills/...).
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to fetch and process arbitrary content from the public internet, creating a significant surface for indirect prompt injection.
  • Ingestion points: Untrusted data enters the agent context through the output of the Tavily and Exa search/extract apps (e.g., tavily/search-assistant, exa/extract) as described in SKILL.md.
  • Boundary markers: Although examples show the use of placeholders like <search-results>, there are no strict instructions or system-level delimiters defined to prevent the agent from obeying instructions embedded within the retrieved web content.
  • Capability inventory: The skill has access to the Bash tool (specifically the belt command) and demonstrates workflows where external data is piped into LLM models for further action.
  • Sanitization: There is no evidence of sanitization, filtering, or validation of the content retrieved from external URLs before it is provided to the agent.
  • [COMMAND_EXECUTION]: The skill relies on the execution of shell commands through the belt CLI. While the allowed-tools configuration limits execution to this specific tool, it represents an reliance on a third-party binary that interacts with external APIs.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 18, 2026, 08:53 PM