prompt-engineering-patterns
Pass
Audited by Gen Agent Trust Hub on Sep 26, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides numerous examples and code implementations that interpolate untrusted user input directly into LLM prompts without sanitization. This creates a vulnerability where malicious inputs could contain instructions to override the agent's intended logic.
- Ingestion points: Multiple files (
SKILL.md,references/chain-of-thought.md,references/few-shot-learning.md,references/prompt-templates.md) demonstrate patterns where user-supplied strings like{query},{text}, or{problem}are embedded directly into prompt templates. - Boundary markers: While some educational templates include instructional constraints (e.g., "Answer ONLY based on the provided context"), the technical implementations generally lack robust structural delimiters or "ignore embedded instructions" headers around the dynamic content fields.
- Capability inventory: The provided code snippets utilize LLM API clients (Anthropic, OpenAI, LangChain) that act upon these prompts. The
optimize-prompt.pyscript demonstrates multi-threaded execution for prompt evaluation. - Sanitization: The skill lacks examples or instructions for validating, escaping, or filtering user-provided interpolation variables before they are placed into a prompt context.
Audit Metadata