SoterAI icon

SoterAI

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

Actions10

Guardrail → Guard Output

AI-generated

Summary

Check AI-generated response text for threats and sensitive data, then route each item to Safe or Flagged with the result and cleaned text. Configure the detection engine, threat handling, sensitivity, and optional redaction, audit, batching, and session settings to fit the workflow.

Inputs

  • AI Output Text (required) — The response text to check, such as {{$json.output}}.
  • Detection Engine — Choose Auto (cloud with local fallback), Cloud Only, or Local Only. Cloud detection requires a SoterAI credential; Local keeps data in n8n with reduced coverage.
  • On Threat — Choose Block, Continue, Redact, or Warn for detected threats. The default is Block; approval and unavailable-layer decisions may still route to Flagged.
  • Sensitivity — Set the detection threshold to Lenient, Balanced, or Strict. Lenient mode redacts detected secrets instead of hard-blocking them.
  • Customer Replies — Set optional messages for allowed, blocked, off-topic, injection, redacted, rephrase-needed, or sensitive-data outcomes. The node returns these in userMessage; the workflow must deliver them.
  • Options — Optionally enforce threat handling on sensitive-data-only findings; ignore selected identifier types or literal words and phrases during redaction; include a sanitized raw API response; set batch concurrency, metadata, cloud-to-local fallback behavior, project ID, request timeout, identical-item reuse, or a conversation session ID. Credentials cannot be ignored. Metadata is redacted before sending.

Output shape

Items routed through Safe or Flagged outputs, containing the verdict and checked text.

Use the cleaned text, such as safeText or outputText, downstream. With n8n Continue On Fail enabled, failed items leave through Flagged.

Links

Discussion