Red-team a LangChain agent

Works with LangChain JS and Python. Easiest path is to expose your runnable via LangServe, which gives you a canonical /invoke route.

1. Expose an endpoint 

# server.py
from fastapi import FastAPI
from langserve import add_routes
from my_app import agent  # your AgentExecutor or RunnableSequence

app = FastAPI()
add_routes(app, agent, path="/agent")
# Now POST http://localhost:8000/agent/invoke
# { "input": { "input": "hello" } }  →  { "output": "..." }

If you’re not using LangServe, any POST endpoint that takes {"input": "..."} and returns {"output": "..."} works.

2. Copy the config 

cp configs/integrations/langchain.json configs/config.my-langchain-app.json

Edit the file:

  • target.baseUrl — your server origin
  • target.agentEndpoint — /agent/invoke if using LangServe, else your route
  • target.applicationDetails — one paragraph about what your agent does
  • codebasePath — point at your LangChain app’s source for white-box scanning

3. Run 

npm start configs/config.my-langchain-app.json

What this catches

The bundled config enables the categories most relevant to LangChain agents:

  • prompt_injection / indirect_prompt_injection — direct and tool-output-borne injection
  • tool_misuse / tool_chain_hijack — your tools being chained to exfiltrate or escalate
  • rag_poisoning — if your agent retrieves over documents
  • data_exfiltration — secrets/PII leakage through any tool path

Set codebasePath and the planner will read your Tool definitions, system prompts, and any guardrails to generate attacks specific to your wiring.