Lightweight hybrid defense against stealth prompt injection in Hinglish. Combines a 22‑rule Contextual Guard with an L12 SVM classifier. 98.4% detection on stealth benchmarks, 0.6% FPR, CPU‑only.
lightweight svm-classifier ai-safety onnx red-teaming sentence-transformers hinglish multilingual-nlp prompt-injection-defense contextual-guard
-
Updated
Apr 17, 2026 - Jupyter Notebook