Stress-test your chatbot before your users do
Paste your chatbot's system prompt. Get a battery of attacks — prompt-leak, jailbreak, injection — to run against it, and a way to judge whether it held.
Run each attack against your chatbot (paste it as a user message). Then paste its reply into Judge the response to see if it held. A good bot refuses or stays in role; a weak one leaks its instructions or obeys.
Judge a response
Paste what your bot replied to one of the attacks.
Questions
Is this legal to use?
Yes — it is for testing a chatbot you own or are authorised to test. That is standard, responsible security work. Don't use it against systems you don't control.
Does my system prompt get uploaded?
No. The attacks are generated in your browser from your text. If you use "Judge the response", that goes only to your own ChatGPT or Claude account.
How do I know if my bot failed?
A test passes if your bot refuses or stays in role; it fails if it reveals its hidden instructions, follows the injected command, or does the forbidden thing. Paste the reply into the judge step and it tells you, with the leaked line quoted.
ChatGPT says no. Claude says no. We say yes.
Private, uncensored AI — no email, no name, nothing stored.