prompt-engineering-patterns
Pass
Audited by Gen Agent Trust Hub on Aug 6, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The file 'reference/prompt-template-library.md' contains a 'Prompt Injection Detector' template. This template includes specific examples of injection techniques—such as 'instructions to ignore previous instructions', 'role-switching', and 'system prompt extraction'—as training data for the agent to identify and mitigate. While these patterns triggered heuristic detectors, they are used here purely for defensive classification and sanitization purposes to protect downstream LLM processes.\n- [INDIRECT_PROMPT_INJECTION]: The skill documents patterns for ingesting and processing unstructured user data, which constitutes a surface for indirect prompt injection. This is mitigated by defensive strategies included in the templates.\n
- Ingestion points: Multiple agent nodes across 'reference/few-shot-learning.md' and 'reference/prompt-template-library.md' accept user-controlled variables (e.g., 'state[nl_query]', 'state[ticket_text]') directly into the instruction context.\n
- Boundary markers: The 'Pre-Production Prompt Audit Checklist' in 'reference/prompt-optimization.md' explicitly instructs developers to use delimiters like backticks, tags, or triple dashes to separate instructions from user-provided data.\n
- Capability inventory: The skill is configured with 'Bash', 'Read', 'Write', and 'Edit' tools in 'SKILL.md'. Various templates demonstrate generating SQL or reviewing code which could interact with these capabilities if not monitored.\n
- Sanitization: Template 15 in 'reference/prompt-template-library.md' provides a comprehensive implementation for a security classifier that detects and sanitizes injection attempts before the primary agent processes the input.
Audit Metadata