shauryasinghh25/sentinel-ai-gateway

SentinelAI Gateway is a security middleware that sits between an application and the large language model providers it calls, such as OpenAI, Anthropic, Google Gemini, or a local Ollama instance. According to the README, every request and response passes through the gateway, where it is checked against a stack of security rules, filtered by configurable policies, and then logged for a real-time dashboard. The gateway is built on FastAPI for the proxy layer, with a React dashboard on the front. It handles JWT authentication and per-key rate limiting, and can fail over to a backup provider if the primary one fails. The injection detection engine ships with what the README calls 50 plus attack patterns, covering categories like jailbreak phrasings, role override attempts, direct instructions to ignore prior context, indirect attacks embedded in retrieved documents, unicode obfuscation, hidden base64 payloads, attempts to misuse tools, and attempts to extract the system prompt. Each check returns a structured score, a risk level, an action, and the matched pattern. There are also dedicated layers for PII detection and redaction across 20 plus categories including emails, phone numbers, API keys for various services, and JWT tokens, plus RAG pipeline checks that scan retrieved chunks for poisoned content before they reach the model. A separate MCP tool security layer enforces allow and deny lists for tool calls and redacts sensitive arguments. Rules are defined in a YAML policy file, with three starter profiles named default, strict, and permissive. The project is meant to be started with Docker Compose, which brings up the API, the dashboard, Prometheus, and Grafana together, with an alternative path that runs the FastAPI backend and the Vite frontend separately for development. Authentication uses an X-API-Key header, and the README lists endpoints for chat proxying, input and output checks, RAG validation, analytics, and policy management. The dashboard shows traffic, blocked attempts, attack type breakdowns, and provider usage. The project is released under the MIT license.