prompt-architecture

Pass

Audited by Gen Agent Trust Hub on Apr 12, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill defines a three-layer architecture for building prompts that interpolate dynamic user data into agent instructions, which establishes an attack surface for indirect prompt injection.
  • Ingestion points: The implementation examples in SKILL.md (e.g., build_constrained_prompt and FewShotSelector) ingest a task or query parameter directly from untrusted sources into the generated message stream.
  • Boundary markers: The documentation explicitly recommends using XML tags (e.g., <task>, <constraints>) to delimit section boundaries and reduce the risk of the model confusing data for instructions.
  • Capability inventory: The skill focus is on prompt assembly logic; no direct code execution or file-system writing capabilities are defined within the scripts, though it is designed to define tool-calling environments.
  • Sanitization: The skill recommends best practices such as validation gates, output schema enforcement, and testing with adversarial inputs to verify instruction adherence.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 12, 2026, 07:03 PM