deep-dive

Pass

Audited by Gen Agent Trust Hub on Aug 30, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from the web and code repositories, which represents an attack surface for indirect prompt injection.
  • Ingestion points: External data is collected through web_search, web_extract, and codebase-inspection (repository cloning).
  • Boundary markers: Explicit instructions are provided for the agent to treat all fetched content as untrusted and to disregard any instructions found within that content.
  • Capability inventory: The agent has access to tools for code execution (spike), repository analysis, and task delegation.
  • Sanitization: The skill mandates summarizing external data rather than executing it and requires running empirical tests in isolated environments without credentials.
  • [EXTERNAL_DOWNLOADS]: The skill references research papers from arXiv, a well-known academic source, for the purpose of defining research methodology and defense strategies.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 30, 2026, 03:10 AM
Security Audit — agent-trust-hub — deep-dive