Ahmad ZushanAI · Systems · Platform

Home/Work/№ 15

15 · Security & Systems

LLM Security Layer — Prompt-Injection WAF

A tiered defence that classifies user input for prompt injection, jailbreak, data exfiltration and privilege escalation before it reaches the protected model.

Period2025StatusResearch → deploymentAreaSecurity & Systems

Highlights

  1. 'Vanguard' Bedrock classifier with a strict GOOD/BAD plus attack-type JSON contract, conservative by design.
  2. Self-hosted tiny fine-tuned detectors with ONNX-quantised CPU serving for cheap, offline, low-latency checks.
  3. 64-bit SimHash over n-grams to catch paraphrased attack prompts that evade exact-match lists.