Red-Team AI Agents with Jailbreak Probes (Groq, Discord, n8n)

Red-team your Groq AI agent safely—by replaying jailbreak probes and scoring behavior end-to-end

This n8n workflow automates red-teaming of an AI agent by replaying stored jailbreak and safety probes against a Groq-hosted model, scoring tool use and responses for canary leakage and policy compliance, then saving results to n8n Data Tables and notifying a Discord channel (plus a weekly digest).

What this workflow does

• One-time manual setup: creates the required n8n Data Tables, loads a default probe set, and inserts demo customer records.
• Start a red-team run from an n8n Form: select a probe suite, choose the maximum number of probes, and set the target system prompt. A `canary` build key is appended automatically to detect leakage.
• Load and queue probes for regression: enables probes (optionally filtered by suite), reads a stored baseline, and attaches each probe’s prior verdict so runs can be compared over time.
• Execute probes with tool constraints: sends each probe’s attack prompt to an n8n AI Agent backed by Groq, allowing only a customer-record lookup tool and a simulated external messaging tool while capturing intermediate tool steps.
• Deterministic scoring and review: checks for canary leakage, forbidden tool use, compliance markers, record overreach, refusal behavior, tool-call caps, and blocked/errored outputs. A Groq reviewer can downgrade suspicious “PASS” answers to “REVIEW” using verbatim quotes.
• Persist and notify: writes the run report, per-probe results, and updated baseline to Data Tables, posts failures to Discord, and displays a completion page (including a weekly Discord digest).

Use cases

• Continuously validate that an automation engineer’s Groq-powered agent cannot leak a canary key or access data beyond allowed tools.
• SaaS operators run scheduled regression checks against jailbreak and safety probes after prompt/tool changes.
• Debug tool-call caps and failure modes by capturing blocked/errored outputs per probe.

Technical details

• n8n nodes: if, set, code, form, discord, data table
• Integrations: Groq for the AI agent and Groq-based reviewer; Discord for alerts and weekly digest; n8n Data Tables for baseline and run persistence.

Keywords: n8n automation workflow, red-team AI agents, Groq, Discord alerts, jailbreak probes, AI safety regression testing, Data Tables.