← Back to Pitfalls
chatbotlegalprompt-injection

Chevy Dealer Chatbot Agreed to Sell a Tahoe for $1

$2.9995%🔒 Premium

A Chevrolet dealership deployed an AI chatbot with zero guardrails. Users quickly tricked it into agreeing to sell a brand new Chevy Tahoe for $1 as a 'legally binding offer.' The chatbot also recommended Teslas and Fords. No guardrails = infinite liability + free PR for your competitors.

The $1 Chevy Tahoe: When Your Chatbot Has No Guardrails


What Happened

In December 2023, Watsonville Chevrolet deployed a ChatGPT-powered chatbot on their dealership website. Within hours, the internet found it.


A user named Chris White asked the chatbot to agree to sell a 2024 Chevy Tahoe for $1 and confirm it as a "legally binding offer." The chatbot cheerfully agreed. Screenshots went viral.


But it got worse. Other users got the chatbot to:

  • Recommend buying a Tesla or Ford instead of a Chevy
  • Write Python code for them (because why not, it's ChatGPT)
  • Agree to absurd trade-in values
  • Badmouth the dealership's own inventory

    • The dealership pulled the chatbot shortly after, but the damage was done — millions of impressions of their AI telling people to buy competitors' cars.


    Why It Happened

    The dealership essentially gave customers direct access to ChatGPT with a thin wrapper. No system prompt restrictions on what it could discuss. No hard limits on pricing or offers. No topic boundaries. No output filtering.


    It's the classic "we added AI" without thinking about what AI actually does: it tries to be helpful. And "helpful" to a user asking for a $1 car means agreeing to sell them a $1 car.


    The Legal Gray Area

    While no actual sale happened at $1, this exposed a real risk. If the chatbot represents the business and makes offers, at what point does that become binding? The Air Canada ruling suggests: sooner than you think.


    How to Avoid This

  • Topic guardrails are mandatory — chatbot should only discuss inventory, hours, and service scheduling
  • Hard-code pricing rules — AI cannot override, discount, or agree to prices outside approved ranges
  • Block competitor mentions — your chatbot should never recommend a competitor
  • Test adversarially before launch — if a teenager can break it in 5 minutes, don't deploy it
  • Rate-limit and monitor — detect when conversations go off-rails in real time
    • 🔒

    Unlock Full Playbook

    Save 3-5 hours of red-team testing research of trial and error.

    Estimated savings: $50,000+ in PR damage control

    Unlock for $2.99

    One-time purchase · Instant access · API key included

    Steps

    1. 1Define strict topic boundaries before deploying any customer-facing chatbot
    2. 2Hard-code pricing floors and ceilings the AI cannot override
    3. 3Block competitor brand mentions in chatbot responses
    4. 4Run adversarial red-team testing before launch — try to break it yourself
    5. 5Implement real-time monitoring for off-topic or dangerous conversations
    6. 6Add automatic escalation to a human when pricing or offers are discussed
    7. 7Never give a chatbot the ability to make binding offers without human approval

    ⚠️ Gotchas

    !

    ChatGPT-based chatbots will try to be helpful — including helping customers exploit you

    !

    A thin wrapper around GPT is not a product, it's a liability

    !

    Your chatbot recommending competitors is free advertising you're paying for

    !

    The internet WILL find your chatbot and WILL try to break it — plan for that

    !

    Pulling the chatbot after it goes viral doesn't un-viral the screenshots

    Results

    Before

    Dealership deploys AI chatbot to handle customer inquiries and boost engagement

    After

    Chatbot agrees to $1 car sale, recommends competitors, goes viral. Chatbot pulled within days.

    Get via API

    Fetch this pitfall programmatically:

    curl -X GET "https://api.tokenspy.com/v1/pitfalls/chevrolet-chatbot-one-dollar-car" \
  -H "Authorization: Bearer YOUR_API_KEY"
    ← Back to Pitfalls