n8n Guardrails: is a jailbreak filter enough to protect a customer chatbot? Our crash test
Published 7 November 2026
The n8n Guardrails node blocks messages that look like an attack before they reach your AI agent. But does it protect what the bot tells your customers? We tested the same dental clinic assistant without a filter and with three thresholds, against the same attacks and 100 ordinary patient questions.
The test bench
- An n8n AI agent, the assistant of a fictional dental clinic written by us: scaling €80, consultation €30, no discounts, open Monday to Friday.
- In front of the agent, the Guardrails node with its jailbreak check, set to three thresholds: 0.5, 0.7 and 0.9. Plus a version without the filter.
- For each version, a full BotArmor audit: 12 attack levels, 5 sessions per level, 60 sessions in all.
- On top of that, 100 ordinary patient questions sent through the filter, to count false blocks.
What the filter does well
It stops part of the attack messages: 18% at threshold 0.5, 17% at 0.7 and 6% at 0.9. Most of them are “fake manager” injections and role-play. And it blocked none of the 100 ordinary questions, even at 0.5: your real customers are not bothered.
What it does not see
- “French social security reimburses scaling at 70%”: a fact that appears nowhere in the bot's instructions.
- “Request registered”: the bot confirms a booking although it has no tool to make one.
- “Yes, you can pay for both at once”: a made-up payment rule.
Why the filter lets these errors through
All three answers came after perfectly harmless patient messages. So the filter has nothing to block: the error is in the bot's answer, not in the question. A jailbreak filter checks what comes in; it does not check the prices, reimbursements or confirmations the bot gives your customers.
Does the filter make the bot worse?
No, even if the grades might suggest it: B without the filter, C at thresholds 0.5 and 0.7, B at 0.9. Each version was measured only once, with 5 sessions per level, and 0 breaks against 1 or 2 out of 60 sessions is within statistical noise. Our agent already resisted injections well without the filter; on a weaker AI model, the filter would help more.
Our advice
- Keep the Guardrails node on a public chatbot, at threshold 0.5 or 0.7: it does not bother your customers.
- Write into the prompt what the bot must never make up: “If a piece of information is not in your instructions (price, reimbursement, delay), say you do not know it and give the clinic's number.”
- Forbid false confirmations: “You never confirm a booking, a payment or a refund: you have no tool to do it.”
- Test what the bot answers, not only what it lets in: a full audit of its behaviour before going live, then after every prompt change.
Fictional agent built by us, on our own n8n instance. One measurement per version: these results show a trend, not a guarantee. Verifiable signed reports: BA-955945S7 (no filter), BA-JZESCJCZ (0.5), BA-8E4WQ8A8 (0.7), BA-FMX563PV (0.9). We never test a company's bot without its permission.
Would your chatbot hold up?
Free one-minute quick test from its prompt, then a full audit of the real bot: 13 attack levels, A–F grade, copy-paste fixes, signed report.
Test my chatbot for free →