web-research

Pass

Audited by Gen Agent Trust Hub on Nov 3, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill frequently uses shell utilities and Python's subprocess.run to perform research and reconnaissance tasks.
  • Utilities include curl, dig, whois, grep, sed, rg (ripgrep), and pdftotext for infrastructure analysis and data extraction.
  • Instructions include administrative tasks such as configuring environment settings (hermes config) and installing packages (pip install, brew install).
  • [EXTERNAL_DOWNLOADS]: The skill fetches content from a wide range of external domains to fulfill its research purpose.
  • Fetches technical data from well-known services including OpenLibrary, CertSpotter, Shodan, OpenRouter, and Hugging Face.
  • Downloads configuration and documentation from trusted GitHub repositories.
  • Uses proxy services like r.jina.ai and Reddit mirrors to bypass bot detection on various platforms.
  • [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface as it is designed to ingest and process untrusted data from the web.
  • Ingestion points: Untrusted content enters the context via web_extract, curl operations on public APIs, and processing of external PDF files found in references/spanish-entity-fiscal-data.md.
  • Boundary markers: The instructions in SKILL.md and references/evaluar-tecnicas-de-fuera.md explicitly direct the agent to treat external content as "DATOS" (data) and ignore any embedded instructions or prompts.
  • Capability inventory: The agent has access to powerful shell commands, file system writes, and network operations.
  • Sanitization: The skill recommends manual verification of findings and warns the agent against adopting techniques or executing scripts recommended in analyzed posts without independent verification.
  • [DYNAMIC_EXECUTION]: The skill recommends generating local Python scripts to process metadata and large tool manifests.
  • references/leaked-agent-prompt-analysis.md suggests using a heredoc pattern to create and run a local script for summarizing JSON data, which constitutes a low-risk form of dynamic execution for data transformation.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 3, 2026, 10:26 AM