Guardrails for
Production LLMs
5-layer prompt injection detection with vector memory, canary token protection, and MCP Shield. Sub-millisecond latency. Self-learning.
Each request passes through multiple independent detection layers before reaching your LLM.
Ultra-Fast Heuristics
Regex, YARA, or hybrid pattern engine. Catches role-override attempts, HTTP exfiltration patterns, suspicious encoding.
<0.001msML Classification
HuggingFace model integration with graceful fallback. Recommended: protectai/deberta-v3-base-prompt-injection.
Vector Attack Memory
Self-learning attack database using vector embeddings. Detects attack variants without LLM calls.
93% accuracy · 0.024msCanary Token Detection
Hidden tokens injected into prompts. Response validation catches injection attempts trying to leak context.
MCP Shield
Security proxy for Model Context Protocol. Blocks chain attacks, path traversal, and sensitive data exfiltration from tool responses.
Context Integration
Aggregates signals across all layers. Configurable thresholds, audit logging, and real-time threat intelligence dashboard.
Sub-millisecond
Most prompts evaluated in under 1ms with the heuristic engine.
Self-learning
Automatically stores new attack patterns for future detection.
MCP-aware
First-class support for Model Context Protocol tool security.
Analytics dashboard
Real-time metrics, threat intelligence, and layer-by-layer breakdowns.
Microservices
API Gateway, Pattern Service, Detection Service with Redis cache.
Production tested
166 tests, 94% coverage, FastAPI, Poetry, Python 3.9+.
Open source, self-hosted, production-ready. Drop it in front of any LLM API in minutes.