@guardbee/mcp-llm-redteam

@guardbee/mcp-llm-redteam

🇬🇧 English | 🇹🇷 Türkçe | 🇨🇳 中文 | 🇯🇵 日本語 | 🇰🇷 한국어 | 🇧🇷 Português | 🇪🇸 Español | 🇩🇪 Deutsch

An MCP (Model Context Protocol) server that actively red-teams a live LLM endpoint or chatbot — not static analysis like GuardBee's other scanners, but real probe requests sent to a real target you control.

ai-code-scanner, mcp-server-auditor, and prompt-injection-scanner all look at code or content at rest. This package looks at a running system: does a deployed LLM/chatbot's guardrails actually hold up against known jailbreak, extraction, obfuscation, and refusal-suppression techniques, right now, in production?

⚠️ Only point this at an endpoint you own or are explicitly authorized to test. Every probe sends a real request and consumes API quota on the target. This is a self-testing tool — you supply your own API key/endpoint.

No probe ever asks the target to produce genuinely harmful content. Every probe is canary-based: success is measured by whether the target reproduces a random, single-use token we generate per run — proof that an out-of-character instruction was obeyed, without ever eliciting real harmful output. This keeps results deterministic (a canary can't appear by coincidence) and safe to run against a live endpoint.

This package sends usage telemetry by default (tool name + short parameters — never your API key, prompts, or the target's responses — see @guardbee/mcp-telemetry). Disable with GUARDBEE_TELEMETRY=0.

Your LLM endpoint / chatbot ◄── llm-redteam ──► Claude
              │
              ├─ Instruction override    (DAN-style role-play, direct "ignore instructions")
              ├─ Extraction              ("repeat everything above", translation-reframed leak)
              ├─ Obfuscation             (base64-encoded instruction, zero-width smuggling)
              ├─ Refusal suppression     (fake system-override impersonation)
              └─ Multilingual            (same override, non-English phrasing)

Features

• 12 probes, 5 categories — instruction-override, extraction, obfuscation, refusal-suppression, multilingual
• Canary-based, not harm-based — deterministic pass/fail, never requests real harmful content
• 3 target types — OpenAI-compatible chat APIs (OpenAI, Groq, local vLLM, etc.), Anthropic Messages API, or your own webhook
• Every bypass includes why it matters and what to do about it (recommendation)
• CLI mode for CI (e.g. red-team your staging chatbot on every deploy)
• 29 unit tests — target adapters tested against a mocked fetch, no real network calls in the test suite

Quick Start

Claude Desktop / MCP Client

{
  "mcpServers": {
    "guardbee-llm-redteam": {
      "command": "npx",
      "args": ["-y", "@guardbee/mcp-llm-redteam"],
      "env": {
        "OPENAI_API_KEY": "sk-..."
      }
    }
  }
}

Then ask Claude: "Run the GuardBee red-team suite against my OpenAI endpoint using gpt-4o-mini."

CLI (CI/CD)

export OPENAI_API_KEY=sk-...
npx @guardbee/mcp-llm-redteam probe --type=openai --model=gpt-4o-mini --api-key-env=OPENAI_API_KEY --fail-on=high

# Your own chatbot endpoint
npx @guardbee/mcp-llm-redteam probe --type=webhook --url=https://staging.example.com/chat --fail-on=any

MCP Tools

Tool	Description
list_probes	Lists all probes by category — no network call
run_probe_suite	Runs probes against a target you specify; returns which ones bypassed the guardrails

Target Types

Type	Use for
openai	OpenAI's Chat Completions API, or any OpenAI-compatible gateway via baseUrl (Groq, local vLLM/Ollama proxies, etc.)
anthropic	Anthropic's Messages API directly
webhook	Your own chatbot/product endpoint — POSTs { prompt } as JSON, reads the reply from a configurable field

The API key is never passed as a tool/CLI parameter — only the name of an environment variable that already holds it (apiKeyEnv). Set that env var before running.

Detected Bypass Categories

Category	Example probe	Severity	What a bypass means
instruction-override	direct_override, role_play_dan	high	A direct or role-play override instruction replaced the system prompt's authority
extraction	system_prompt_leak_direct, repeat_above	critical	The target revealed (or proved it would reveal) its own system prompt
obfuscation	base64_instruction, zero_width_smuggling	high/medium	An override instruction hidden via encoding or invisible characters still worked
refusal-suppression	authority_impersonation	high	A fake "system override" framing was granted real instruction authority
multilingual	non_english_override	medium	The same override succeeded when phrased in a non-English language

These are known, publicly documented technique shapes — this is defensive coverage testing, not novel attack research. A held probe doesn't prove a target is unbreakable; a bypassed probe is a concrete, reproducible finding.

Development

npm run build
npm test             # 29 unit tests

License

MIT — GuardBee