prompt-caching-patterns

Pass

Audited by Gen Agent Trust Hub on Aug 31, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill contains architectural guidance and code snippets for implementing prompt caching. All identified external dependencies and service references (Anthropic, OpenAI, Redis) are well-known and trusted technology providers. The implementation logic follows industry standards for hashing and data storage.
  • [INDIRECT_PROMPT_INJECTION]: The skill provides patterns for caching user prompts and LLM responses, which inherently involves processing untrusted data. While this creates a potential surface for indirect injection if cached content is not properly isolated or validated, the provided examples are educational in nature and follow standard software development practices.
  • Ingestion points: User-supplied prompts are processed in functions such as queryWithCache, cachedQuery, semanticCache.get, and templateCache.get in SKILL.md.
  • Boundary markers: The code templates do not demonstrate the use of explicit delimiters or boundary markers to separate instructions from data within the cached prompts.
  • Capability inventory: The code includes capabilities for performing network requests to LLM APIs and communicating with external Redis storage for cache persistence.
  • Sanitization: The provided examples focus on the caching mechanics and do not include specific input sanitization or output validation steps for the cached content.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 31, 2026, 01:19 PM
Security Audit — agent-trust-hub — prompt-caching-patterns