Langgraph Guard

This package is required for integrating your Langchain based Agentic Application to Aryaka's AISecure Guard Service.

What it does

langgraph-guard is a security middleware designed to seamlessly integrate with any LangGraph application, providing a critical layer of real-time security/content validation. It intercepts user inputs, LLM prompts and responses, tool I/O, retriever I/O, and final outputs via a LangGraph callback handler and asynchronously calls the AISecure GenAI Protect service. If this service identifies a policy violation, langgraph-guard immediately blocks the unsafe operation from continuing, preventing threats before they can impact the application.

Configuration is managed declaratively through environment variables, requiring no changes to the application's source code.

Key Capabilities

• Real-Time Threat Prevention: Actively blocks malicious inputs, prevents sensitive data (PII) leakage, and ensures AI-generated content adheres to organizational safety standards.
• Context-Aware Policies: Applies granular security rules based on the specific user, their role, and the task being performed within the AI system.
• Comprehensive Validation: Leverages a powerful suite of validators including but not limited to:
• Content & Safety: Toxicity, sentiment, and content classification.
• Privacy & Compliance (DLP): PII detection and adherence to data protection standards.
• Code Detection & Prompt Injection: Detection of code snippets and potential prompt injection attacks.
• Zero-Code Integration: A fast, non-invasive way to add robust security and compliance controls to any LangGraph-based application

Configuration (via environment variables)

All configuration is loaded from environment variables (e.g., from a .env file). At minimum, enable the guard and provide the service URL, default inspection object, stage mapping, and tenant/site identifiers.

Required when GUARD_ENABLED=true:

• GUARD_URL: Base URL of the AI Secure Validation API.
• GUARD_DEFAULT_INSPECT: Fallback inspection object name.
• GUARD_STAGE_MAP: JSON mapping that selects inspection objects by stage/group/user/hook.
• GUARD_CUSTOMER_ID: Customer ID
• GUARD_TENANT_ID: Tenant ID.
• GUARD_SITE_ID: Site ID.
• TLS either:
  • GUARD_INSECURE_SKIP_VERIFY=true (dev only), or
  • both GUARD_CA_PATH and GUARD_CA_PEM set for certificate verification.

Optional:

• GUARD_JWT: Bearer token for the guard service.
• GUARD_USER_ID: Default user id (if not provided at runtime).
• GUARD_GROUPS: JSON array of groups for policy resolution (e.g., ["analysts","admins"]).

Example .env snippet:

GUARD_ENABLED=true
GUARD_URL=https://protect.example.com
GUARD_JWT=eyJhbGciOiJI... # optional
GUARD_DEFAULT_INSPECT=inspect_default
GUARD_STAGE_MAP={
  "stage:plan": {"pre_llm": "inspect_plan"},
  "group:analysts": {"pre_llm": "inspect_group"},
  "user:alice": {"final_output": "inspect_user"},
  "*": {"final_output": "inspect_default"}
}
GUARD_CUSTOMER_ID=ciid-123
GUARD_TENANT_ID=tenant-abc
GUARD_SITE_ID=site-001
# One of the following TLS setups
GUARD_INSECURE_SKIP_VERIFY=false
GUARD_CA_PATH=/etc/ssl/certs/ca-bundle.crt
GUARD_CA_PEM="-----BEGIN CERTIFICATE-----\n...\n-----END CERTIFICATE-----\n"

GUARD_STAGE_MAP shape (keys → hook → inspect name):

• Keys may be:
  • stage:<name> or just <name> for a workflow stage
  • group:<name> for group-based rule
  • user:<id> for user-specific rule
  • * wildcard for any stage
• Hooks supported by the handler:
  • user_input, final_output, pre_llm, post_llm, pre_tool, post_tool, pre_mcp, post_mcp, error

How to use

You can attach the guard with zero code changes using environment variables, or explicitly in code.

1) Environment-driven (no code changes)

Wrap your LangGraph runnable factory with the provided decorator. The guard will read all settings from env and attach itself only if GUARD_ENABLED=true and required vars are present.

from langgraph_guard import guard

@guard
def get_app():
    # build and return your LangGraph runnable
    return app

Alternatively, if you already have an app instance:

from langgraph_guard import attach_guard_from_env

app = attach_guard_from_env(app)

2) Programmatic attach (custom config objects)

If you centralize config in your app, you can pass that object to attach based on its attributes:

from langgraph_guard import attach_guard_if_enabled

class AppConfig:
    guard_enabled = True
    guard_url = "https://protect.example.com"
    guard_jwt = "..."  # optional
    guard_default_inspect = "inspect_default"
    guard_customer_id = "ciid-123"
    guard_tenant_id = "tenant-abc"
    guard_site_id = "site-001"
    guard_ca_path = "/etc/ssl/certs/ca-bundle.crt"
    guard_ca_pem = None
    guard_insecure_skip_verify = False
    # Optional identity defaults
    guard_user_id = "anonymous"
    guard_groups = ["analysts"]

# Ensure GUARD_STAGE_MAP is provided via environment
app = attach_guard_if_enabled(app, AppConfig())