Dicklesworthstone/acip

Deploys a multi-layered prompt injection defense by combining explicit narrative directives with categorized attack examples, guiding models to detect semantic manipulation, obfuscation, and indirect injection via RAG/tools. Available in three versioned variants (v1.0–v1.3) with trade-offs between token overhead, domain coverage, and operator observability—v1.3 adds configurable audit mode for security monitoring while v1.2 prioritizes cyber hardening. Integrates directly into LLM workflows via prepended prompts, dedicated checker models, or hybrid approaches with minimal architectural changes.

300 stars.