Skip to main content
ShieldPrompt is the orchestration pipeline that combines four GPU-powered model-based detection layers in defence-in-depth. Each layer catches a class of attack the others miss; run together they compose into resilient runtime defence.

​
LLM Classifier

An LLM-as-judge classifier that scores prompt-injection likelihood and threat intent on prompts and outputs. Thresholds are configurable per security profile; defaults are tuned for production.

​
Vector Filter (TVDB)

Semantic-similarity search against CloudsineAI’s proprietary Threat Vector Database. Catches paraphrased and rephrased injection attempts that bypass keyword and regex filters.
  • Sensitivity: Low / Medium / High — selectable per profile.
  • Catches semantic variants of attacks already in the TVDB; resilient to surface-level rewording.

​
Content Moderation

Purpose-built content-moderation classifier covering hazardous-content categories (violence, hate speech, unethical instructions, sexual content, and similar). The response includes a category label when content is unsafe:

​
System Prompt Protection

LLM-based detection of system-prompt leakage in LLM responses. No application-side configuration is required.

​
Why four layers, not one

Each layer catches what the others miss. A classifier alone misses semantically novel attacks; vector similarity alone misses entirely new attack classes; content moderation alone misses the prompt-injection vector; system-prompt protection alone misses everything that doesn’t leak the prompt. Run together, the layers compose into defence-in-depth.