SoterAI icon

SoterAI

Detect prompt injection, jailbreaks, secrets, PII, and unsafe AI instructions in n8n workflows

Actions10

Guardrail → Universal Firewall

AI-generated

Summary

Checks user messages and AI responses for threats like prompt injection and sensitive data using a cloud or local engine, and optionally validates RAG contexts, tool calls, and memory operations.

Inputs

  • inputText — The user message to check for prompt injection, jailbreaks, and other threats. Optional when AI Output Text is supplied.
  • universalOutputText — Optional AI response to check before sending, saving, or calling another tool. Optional when Input Text is supplied.
  • detectionEngine — Where checks run. Cloud requires a SoterAI credential; Local keeps data inside n8n.
  • protectionProfile — How strict detection is. Maximum suits production and public chatbots.
  • onThreat — How to handle a detected threat. Block sends threats to Flagged. Approval and unavailable-layer decisions may still route to Flagged.
  • securityContext — Add the RAG, tool, memory, or output destination your workflow uses. Each layer is optional.
  • userMessages — Optional customer-facing replies. Results are returned in userMessage; your workflow delivers them.
  • options — Optional detection, audit, and execution settings. Defaults apply until a setting is added.

Output shape

a list of items, with flags, findings, and sanitized text in the JSON data, and optional customer replies in a userMessage field

Items are routed to Safe or Flagged based on threat verdict. The Flagged output also includes details of the detected issue. When onThreat is set to Redact or Warn, the original item is returned with sanitized text. Reused items are marked reusedResult.

Examples

Example 1: Blocking a malicious prompt injection attempt

Set Detection Engine to Auto or Cloud, Protection Profile to Maximum, and On Threat to Block. Input text contains the malicious prompt.

Example 2: Checking an AI response before sending to a user

Supply AI Output Text and set Output Destination to Final Output (Back to the User). On Threat can be set to Redact to remove sensitive data.

Example 3: Validating RAG context for poisoning

Add a Retrieved Context (RAG) layer with the document text and source type. On Threat defaults to Block to prevent contaminated information from flowing downstream.

Links

Discussion