Enterprise LLM Security Firewall and Guardrails - n8n Workflow

• AI Infrastructure and DevOps teams requiring input/output validation for LLMs.


• Security professionals focused on mitigating risks associated with Generative AI deployment.


• n8n specialists looking for advanced n8n templates for AI safety.


• Developers building secure, production-ready AI agents.

When deploying large language models (LLMs) in an enterprise environment, security and compliance are paramount. Unrestricted access can lead to data leaks, prompt injection (jailbreaking), and generation of harmful or off-topic content. This robust n8n workflow serves as an AI security firewall, ensuring every prompt passes rigorous checks before interacting with the LLM, or before its output is returned to the user.

This specific n8n workflow uses a suite of specialized n8n node actions (Guardrails) powered by a foundational LLM (Google Gemini) to categorize and mitigate risks related to PII, secret keys, NSFW content, and topical misalignment. Utilizing this n8n template helps you maintain control, ensuring your AI initiatives are safe, predictable, and compliant.

The entire process is initiated by a manual n8n trigger for testing and demonstration purposes.

1. Data Input: The workflow starts with the When clicking ‘Execute workflow’ n8n trigger, which immediately fetches test data (simulated prompts) from the Guadrails Test Data Google Sheets n8n node.


2. Routing Inputs: The input data is passed to a Switch n8n node, which routes the input message to the appropriate guardrail n8n node for inspection.


3. Security Checks: The prompt is simultaneously evaluated by seven distinct Guardrail n8n node types:

1. Sanitization Pathways: If specific risks are detected (PII, Secret Keys, or URLs), the flow branches to dedicated sanitization modules (Sanitize PII, Sanitize Keys, Sanitize URLs). These guardrail n8n nodes attempt to mask or remove the sensitive data, providing secure output, demonstrating the comprehensive capabilities of this n8n workflow structure.

To deploy this enterprise-grade n8n workflow and start using these guardrail checks, follow these steps:

1. Import the n8n Workflow: Copy the provided JSON data and import it into your n8n instance via the 'Workflows' menu using the 'New' button and selecting 'Import from JSON'.


2. Configure Credentials:

1. Review Guardrail Configuration: Open each Guardrail n8n node (e.g., Keywords, PII) and adjust the internal configuration (like allowed topics or blocklists) to match your organization's specific security policies. This is essential for customizing the n8n templates to your environment.


2. Testing: Execute the workflow manually using the When clicking ‘Execute workflow’ n8n trigger to observe how prompts are routed and flagged by the security checks.

When clicking ‘Execute workflow’ (Manual Trigger): Acts as the entry point, allowing users to start this n8n workflow on demand for testing.
Guadrails Test Data (Google Sheets): This n8n node retrieves the sample prompts that will be subjected to the various security checks implemented by the n8n workflow.
Switch n8n node: Routes the incoming test data to the appropriate security check guardrail based on criteria (though in this demonstration, it connects all paths for simultaneous testing).
Google Gemini Chat Model (AI Language Model): Provides the underlying large language model engine used by all Guardrail n8n nodes to assess, score, and potentially sanitize the input and output data within this n8n template.
Guardrails n8n node (Jailbreak, PII, NSFW, etc.): These specialized n8n node types are the core security mechanism, defining the rules and constraints for safe interaction with the LLM. They determine if the input passes the security protocols set in this n8n workflow.
Sanitize PII, Sanitize Keys, Sanitize URLs (Guardrails and Set nodes): These nodes demonstrate post-detection actions. The 'Set' nodes (Personal Data, API Keys, URLS) prepare the data for the final Guardrail n8n node, which performs the actual redacting or masking of sensitive information.