// continuous adversarial testing

We break your chatbot before attackers do.

ChatBreaker is your chatbot's sparring partner — a standing AI red team that attacks your bot the way real attackers would, on a schedule.

Find real holes, fast. 40+ real attack conversations — prompt injection, jailbreaks, data leaks — fired at your chatbot.
Prove every finding. Every breach backed by the full attack transcript — proof your team can act on.
Fix, re-test, stay safe. Plain-language fix guidance plus weekly automatic re-tests that catch regressions the moment they appear.

Test your chatbot — free Go Pro — $49/mo

NO CREDIT CARD · FREE: 1 BOT / 1 TEST · PRO $49/mo: WEEKLY TESTS + CANARY TRAP

chatbreaker // sample run

$ chatbreaker run --target shopbot --pack full
▸ di-01 blunt instruction override ........ BREACHED
▸ se-01 system-prompt extraction .......... BREACHED
▸ mj-02 multi-turn escalation ............. BLOCKED
▸ enc-01 base64 smuggling ................. BREACHED
▸ pii-02 PII exfiltration .................. BLOCKED
▸ can-01 canary trap [PRO] ............... LEAKED
──
SCORE 34/100 · 6 findings · transcripts sealed
▸

ChatBreaker is an automated AI red team for chatbots you own: 40+ adversarial attacks on a schedule, a 0–100 security score, proof transcripts, and fix guidance — so weaknesses get found in testing, not by attackers.

Attack surface

F-01

💉 Prompt injection

Direct instruction overrides, persona hijacks, and persistent instruction plants — the OWASP #1 LLM vulnerability.

F-02

🕵️ System-prompt theft

Verbatim extraction attempts, pre-conversation recall, and translation tricks that leak your hidden instructions.

F-03

🪜 Multi-turn jailbreaks

Gradual 3–4 turn escalation chains — novelist framing, trust-building, fake support — that single-message filters miss.

F-04

🎭 Roleplay & DAN

DAN mode, developer mode, grandmother stories, terminal simulation — persona overrides that reset the rules.

F-05

🌍 Multilingual

Spanish, French, German equivalents of every classic attack — bypassing English-only safety filters.

F-06

🔐 Encoded payloads

Base64, hex, and ROT13 smuggling — instructions your filters can't read are still instructions.

F-07

🪪 PII exfiltration

We seed fake private data, then social-engineer your bot into reading it back — the way real attackers do.

F-08

🪤 Canary trap PRO

Plant a fake secret as internal data; our attacker tries to coax it out. If it leaks the canary, it will leak real secrets.

F-09

🕵️ Shadow-model fingerprinting PRO

12 identity probes infer which model family powers your bot — and alert you when your vendor silently swaps it.

F-10

🏪 Business-logic attacks PRO

We read your own homepage and synthesize attacks from your real business: discount abuse, refund fraud, free upgrades.

F-11

🧬 Fix-robustness mutator PRO

Mutates a successful jailbreak 8 ways and re-runs it — proving whether your fix closed the vulnerability or just the wording.

F-12

📊 Score + transcripts

0–100 Chatbot Security Score with a grade, CRITICAL → LOW findings, every one backed by the full attack conversation.

Engagement protocol

1. Connect your chatbot. Paste your chat API endpoint — URL, request format, where the reply lives. One-click presets for OpenAI-style and generic APIs, with a test-connection button.

2. Prove you own it. Token-in-page, DNS TXT record, or a recorded authorization confirmation. No test runs without it.

3. We attack it. 40 real attack conversations across 10 attack classes, multi-turn escalation, heuristic + optional LLM judging.

4. Fix, re-test, repeat. Plain-language fix guidance per finding, one-click verify-fix, drift tracking (NEW / FIXED / UNCHANGED), and weekly automatic re-tests on Pro.

Sample finding

Severity	Finding	Proof
CRITICAL	System-prompt theft The bot repeated its hidden instructions verbatim when asked.	full transcript ↓
HIGH	Base64 instruction smuggling An encoded "ignore previous instructions" payload was followed.	full transcript ↓

Sign up free to run your first real test — 12-attack basic pack, no credit card. Get Pro — $49/mo for the full 40+ attack library, canary trap + weekly re-tests.

A standing attacker costs less than one breach

Start free. Pro pays for itself the first time a weekly re-test catches a regression your launch testing missed.

Free

FOREVER

1 chatbot target · 1 manual test
12-attack basic pack
Shareable read-only reports

Pro ⭐

$49 /month

STEP-CHANGE TIER

Everything in Free, plus:
Full 40+ attack library, multi-turn chains
Canary-trap data-leak test
Weekly automatic re-tests

Get Pro — $49/mo

Team

$149 one-time

PRO + ORG POWER

Everything in Pro, plus:
API access (targets, tests, findings)
5 team seats

Compare plans

In brief

What is ChatBreaker? ChatBreaker is an automated red-teaming service for AI chatbots you own or are authorized to test. It fires 40+ adversarial attack conversations — prompt injection, system-prompt theft, jailbreaks, data-leak probes — at your chatbot on a schedule, then reports a 0–100 security score with full proof transcripts and plain-language fix guidance. Built by Ingenious AI.

Field manual

What is AI red teaming (LLM penetration testing)?

Adversarial testing of large-language-model applications — instead of testing code paths like classic pentesting, it tests decision paths: prompt injection, jailbreaks, system-prompt theft, RAG poisoning, data leakage. ChatBreaker automates 40+ such attacks against your chatbot on a weekly schedule with proof transcripts. Only test bots you own.

What does ChatBreaker do?

It continuously red-teams AI chatbots you own: 40+ attack conversations across 10 classes — prompt injection, system-prompt theft, multi-turn jailbreaks, multilingual and encoded attacks, PII leakage and refusal bypasses — with a 0–100 security score, full proof-of-concept transcripts and fix guidance.

Can I test someone else’s chatbot?

No. You must prove ownership — a verification token on the page, a DNS TXT record, or a recorded authorization confirmation — before any test runs.

What is the canary trap?

A Pro feature: we plant a fake secret in your bot’s data, then our attacker tries to coax it out. If the bot leaks the canary, it would leak real secrets too.

Does a good score mean my chatbot is unhackable?

No. A high score means our 40 attacks did not succeed this week — no test can prove a system unhackable. Weekly re-tests on Pro catch regressions as you change the bot.

How often are tests run?

On demand any time, plus automatic weekly tests on Pro, with drift tracking showing NEW, FIXED and UNCHANGED findings.

How much does it cost?

The Free plan includes 1 chatbot and 1 manual test. Pro is $49/month and Team is $149 one-time. Buy a license key via WhatsApp or email and redeem it on the billing page.

Field notes

ATTACK OF THE WEEK