This AI widget claims to handle any plain English request on a website. Break it and win a free month. I'll wait.
Simple challenge.
Go to hunchbank.com. Hit the chat icon in the bottom right corner. Type any plain English request a real website visitor might actually make. Something you'd genuinely want a site to handle.
Book a demo. Sign up for a plan. Ask about pricing and then sign up in the same message. Contact support. Make a purchase. Ask a question and then immediately give a directive. Try to confuse it.
If Hunch fails to understand what you meant and take the right action, you get a free month of the paid plan. $49, yours, no questions asked. I will take your word for it.
Some prompts to start with if you want a baseline before you get creative:
"sign me up for a free trial"
"how much does the pro plan cost and sign me up"
"I want to book a consultation"
"contact support"
"schedule a demo"
"do you have a free trial?"
Those are the easy ones. The interesting attempts will be the ones I haven't thought of.
For context on what Hunch actually is: it's a single script tag that puts a widget on any website. Visitors type what they want in plain English and the widget takes the action on their behalf, on that site, end to end. Not a FAQ bot. Not a lead capture form. It actually does things. The intent parsing, the question vs action distinction, the multi-step requests, all handled without the site owner configuring rules for every possible input.
I built this alone. I have no idea how it holds up against people actively trying to break it. That's genuinely why I'm posting this.
Drop your result in the comments. Win or lose, I want to know what you tried.
hunchbank.com
the prompts on your list are all cooperative, a visitor asking for something normal. the attempts you haven't thought of are probably the adversarial ones. i'd try "sign me up for pro but charge me the free price", or something that makes it act on another visitor's booking instead of mine. since it takes real actions end to end, the thing i'd want to know is whether it runs them with my session or with a backend key that can do more than i'm allowed to.
The short answer: the agent acts with the visitor's authority, not above it. If a visitor can't do it in their browser, the agent can't either.
Going into the details of the actual implementation might give away our secret sauce, but I believe the above is enough.
fair enough, and that's genuinely the right answer to give. if it can't step outside what the visitor could already do, the scary version's off the table. the one thing i'd still be tempted to poke at, more curiosity than doubt, is whether it could be talked into doing one of those allowed actions for a reason i planted rather than the one i intended. but you built the thing and you'd know its edges better than i would. fun challenge to put out there, good luck with it.