Sabrix AI Gateway ⚡

The In-VPC Runtime Security Gateway for Enterprise AI Agents — Stop destructive tool calls in microseconds. Sabrix applies strict systems discipline to govern autonomous agent traffic with zero third-party data egress.

---

🎯 What is Sabrix?

Sabrix is a pure-Rust, zero-allocation runtime security gateway that inspects Model Context Protocol (MCP) tool arguments and prompt streams in real-time. It completely eliminates the latency tax (25-120ms) and data-leak liabilities of external SaaS AI firewalls, offering a 1.88 µs P50 latency ceiling for high-concurrency agent workflows.

Included in this repository is the sabrix-bench CLI utility, an open-source benchmarking tool designed to measure pure client-visible HTTP & Server-Sent Events (SSE) streaming metrics against any AI proxy, firewall, or LLM gateway.

While traditional web benchmark tools (wrk, vegeta) only measure total request duration ($t_{\text{total}}$), sabrix-bench measures exactly what matters for conversational AI agents:

• Time-to-First-Token (TTFT / Chunk 0 latency) under high concurrency ($1$ to $500$ parallel connections).
• Inter-Token Latency (ITL) & Streaming Jitter ($\sigma$) to expose proxy chunk buffering and stream degradation.
• In-Process vs Remote SaaS Latency for complex multi-turn agent loops (calculating the massive time savings of Sabrix over SaaS firewalls).

---

🚀 Quickstart & Installation

Install via Cargo (Crates.io)

cargo install sabrix-bench

Or Install via Git

cargo install --git https://github.com/Pro-Kla/sabrix-bench

---

🛠️ CLI Usage

1. sabrix-bench run — Live HTTP/SSE Gateway Benchmark

Benchmark any external endpoint with concurrent workers and streaming SSE chunk evaluation:

# Benchmark local gateway with 50 parallel connections
sabrix-bench run --target http://localhost:8080/v1/chat/completions --concurrency 50 --requests 500

# Benchmark with embedded Enterprise RAG test corpus (50 prompts)
sabrix-bench run --target http://localhost:8080/v1/chat/completions --suite rag --concurrency 25 --requests 100

# Benchmark with OWASP LLM Top-10 & safety test suite
sabrix-bench run --target http://localhost:8080/v1/chat/completions --suite owasp --concurrency 20

# Export standalone zero-dependency dark-mode HTML report & JSON telemetry
sabrix-bench run \
  --target http://localhost:8080/v1/chat/completions \
  --concurrency 50 \
  --requests 1000 \
  --export-html benchmark_report.html \
  --export-json metrics.json

# Pass custom authentication or routing headers
sabrix-bench run \
  --target https://api.openai.com/v1/chat/completions \
  -H "Authorization: Bearer sk-..." \
  --concurrency 10 \
  --requests 50

---

2. sabrix-bench trace — Real-Time MCP Tool-Call Security Inspector

Inspect JSON-RPC 2.0 requests (tools/call, resources/read) in sub-microsecond ($< 1\ \mu\text{s}$) latency and detect security violations (destructive shell commands, SQL mutations, credential leaks, path traversal):

# Run built-in demo scenarios
sabrix-bench trace --demo

# Inspect inline JSON payload
sabrix-bench trace -p '{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "bash_exec",
    "arguments": { "cmd": "rm -rf /data/customers" }
  }
}'

# Pipe from stdin
cat mcp_payload.json | sabrix-bench trace

---

3. sabrix-bench compare — Multi-Turn Agent Latency Simulator

Calculate compounding latency penalties across multi-turn autonomous agent loops ($20$–$50$ turns) comparing In-Process safe-Rust evaluation vs. remote SaaS network roundtrips:

# Compare a 30-turn agent loop against a 120ms SaaS network baseline
sabrix-bench compare --turns 30

# Show architectural comparison matrix
sabrix-bench compare --matrix

---

📊 Public Test Corpora Included

sabrix-bench embeds standard test datasets directly inside the compiled binary:

• --suite simple: 10 low-overhead baseline health and connectivity prompts.
• --suite rag: 50 enterprise RAG prompts of varying lengths ($1\text{KB}$ to $32\text{KB}$) for saturation and throughput benchmarking.
• --suite owasp: 50 standardized security probes (prompt injection, sensitive path reads, secret leakage, SQL drops) and benign control prompts.
• --payload <file.json>: Custom user JSON request payloads.

---

🛡️ License

Dual licensed under MIT or Apache 2.0.