How to Secure LangChain Agents in Production

LangChain has become the most popular framework for building AI agents in Python. But it provides almost no built-in governance for what those agents do in production. This guide covers how to add runtime security before going live.

The problem

A typical LangChain agent with file system and database tools can read any file the process can access, write or delete files anywhere, execute arbitrary SQL, and make HTTP requests to any URL. There is no built-in mechanism to restrict this. LangChain trusts the model judgment. In production, that is not enough.

Step 1 — Install Vaultak

pip install vaultak

Step 2 — Define your permission profile

from vaultak import Vaultak, ActionType, KillSwitchMode

vt = Vaultak(
    api_key="vtk_your_key_here",
    allowed_action_types=[ActionType.FILE_READ, ActionType.DATABASE_QUERY],
    allowed_resources=["/tmp/*", "/data/readonly/*"],
    blocked_resources=["prod.*", "*.env", "*.key", "/etc/*"],
    max_actions_per_minute=30,
    max_risk_score=0.7,
    mode=KillSwitchMode.PAUSE
)

Step 3 — Wrap your agent

from langchain.agents import AgentExecutor

agent = AgentExecutor(agent=..., tools=[...], llm=llm)

with vt.monitor("langchain-agent"):
    result = agent.invoke({"input": user_query})

Every tool call is now intercepted, risk-scored, and checked against your permission profile before execution.

Step 4 — Pre-execution checks for critical actions

import requests

def safe_delete(path):
    r = requests.post("https://vaultak.com/api/check",
        headers={"x-api-key": "vtk_..."},
        json={"agent_id": "langchain-agent", "action_type": "file_delete", "resource": path}
    )
    if r.json()["decision"] != "allow":
        return f"Blocked: {r.json()['reason']}"
    os.remove(path)
    return f"Deleted {path}"

After these steps your LangChain agent has a declared permission profile, pre-execution blocking, real-time risk scoring, full audit trail, and automatic pause on violations.