Key Takeaways

  • Claire Vo lost tens of thousands of dollars in compute costs over a single year at ChatPRD due to automated token fraud.
  • Credit card verification fails to stop token theft because attackers use automated card cycling, stolen cards, and chargebacks to burn through model access.
  • Attackers exploit every zero-dollar gateway, including promo code reselling rings and malicious Model Context Protocol (MCP) servers configured to drain downstream API calls.
  • Defending AI margins carries a direct penalty on acquisition, because every defensive filter you add creates friction for paying buyers.

Why Credit Card Checks Fail Against Compute Arbitrage

For fifteen years, software founders relied on a simple rule: give away product access for free, let users build habits, and convert them to paid tiers later. If a free trial user logged in every day, your hosting bill barely noticed. Database writes and static assets cost fractions of a cent.

Vo learned that this playbook collapses when you sell models. When Vo ran ChatPRD with trial access, attackers treated her interface as an unpaid proxy for expensive model completions. “Every zero dollar door gets used, every trial, every coupon code. There is no free lunch anymore. It's not like seats, it's not like when your hosting would go up a little,” (play from 35:48) Vo explained.

When Vo added credit card requirements to stop bot farms, the attacks adapted instead of stopping. Fraud rings automated card cycling, testing batches of stolen or virtual cards against her checkout flow to open fresh trials. Each successful card gave the bot farm an active token session to generate completions, leaving Vo to foot the API bill and deal with bank chargeback fees weeks later. As Vo put it, “We had trials require credit card and they would credit card cycle, use bad credit cards, do charge backs. And so actually credit card backed trials and even paid trials don't work.” (play from 15:25)

Protecting Margin Always Hurts Your Real Funnel

When software margins were 85%, you could treat fraud as a rounding error. When you sell model calls, you are reselling raw compute from providers like OpenAI and Anthropic. A single malicious user running batch scripts against your endpoints can rack up thousands of dollars in token usage in one night.

Vo watched attackers exploit promo codes by reselling them on third-party forums, turning ChatPRD promotional budgets into someone else's arbitrage business. Attackers also plugged malicious custom MCP servers into the application to extract underlying model outputs without paying retail rates. Vo fought back by writing strict business rules, rate limits, and automated verification workflows to flag suspicious usage patterns.

The real cost of those defenses was not just engineering hours. Every barrier erected against fraudsters slowed down genuine product managers trying to test ChatPRD for work. “The problem with this, the thing that sucks so much about this is every time I do something that protects us from fraud, it has a negative impact on our real business and our real prospects,” (play from 6:30) Vo noted. Her conclusion was clear: “The first thing is I think free trials over. I'm calling it right now. It's just not worth it, man. I am going to shut the door.” (play from 15:10)

What to Do With This

Audit your checkout flow this week and remove every zero-dollar entry point, including card-backed free trials and unrestricted promo codes. Replace your trial with an upfront paid tier, a strict usage cap tied to a hard balance, or an interactive product demo that generates pre-computed outputs instead of live API calls.