context-engineering

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The multi-agent workflow prompt templates interpolate dynamic external data such as agent outputs and claims, which introduces an indirect prompt injection surface vector.
  • Ingestion points: Dynamic variables like {agent_output} and {claim} are used within prompt blocks in SKILL.md to process external outputs.
  • Boundary markers: The prompt templates utilize structural XML elements (e.g., <OUTPUT_TO_ANALYZE>, <CLAIM>) to encapsulate the interpolated data, which helps mitigate accidental interpretation of data as instructions.
  • Capability inventory: The documentation references the usage of file navigation and reading tools, but no executable code or active tool definitions are supplied within the skill itself.
  • Sanitization: The templates do not explicitly mention or demonstrate escaping methods (such as sanitizing nested XML tags) for the interpolated variables.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 08:31 AM
Security Audit — agent-trust-hub — context-engineering