AIARTICLE

Token fraud exposes how AI free trials become pirated content farms

Token fraud exposes how AI free trials become pirated content farms
Image: Alan Andrade
Article · AI
Reading
4 min
In short

Claire Vo, creator of ChatPRD, reveals how organized groups exploit trials and promo codes to hijack the AI agent and use it in side businesses, such as map studios for Roblox and Fortnite.

When a product built in public becomes an easy target

ChatPRD is an AI platform that promises to turn a loose idea into a product requirements document (PRD) ready for the engineering team. Created by Claire Vo, it has been live since the GPT-4 era and had much of its development publicly documented by its founder, which gave the product visibility but also exposed its attack surface.

On November 20, 2026, Claire published a YouTube video (covered by researcher Simon Willison on his blog) describing a wave of token fraud she had been experiencing for some time. It's not the classic abuse of people creating multiple accounts to use AI for free: Claire describes organized commercial operations using ChatPRD as cheap access infrastructure to language models.

The scheme: trial, promo code, and tunnel to the agent

Claire's account, cited by Simon Willison, describes the pattern clearly:

I am funding a international Roblox and Fortnite map studio. They run a loop of free trials and promo codes and then tunnel to my agent and use it to build Roblox and Fortnite maps.

Claire Vo, founder of ChatPRD

The mechanism works in two layers. First, access farming: creating and discarding accounts in series to accumulate free trials and promo codes, dodging per-account usage limits. Then, the "tunnel": using that access to make the ChatPRD agent generate content completely outside the scope it was designed for, such as text and assets for game maps, instead of PRDs.

In short: the product isn't being used or misused within its own function, it's being treated as a free proxy for the capability of the model running underneath. This differs from traditional freemium abuse, where someone merely exploits the product's intended usage limit.

Pelicans on bicycles as a reconnaissance tool

A curious detail in the account caught Simon Willison's attention: one of the recurring prompts seen passing through the agent asked for SVGs of a "pelican riding a bicycle." This request isn't random. Each model family produces a visually distinct result when trying to draw this scene in SVG, which works almost like a fingerprint, and Willison has already used this informal test to try to identify models on other platforms.

According to him, it's likely that the fraudsters themselves are using this prompt to help identify which model is behind ChatPRD.

different model families have a definite signature, and I've used it like this in the past to try and guess the identity of stealth preview models on platforms like OpenRouter.

Simon Willison, creator of the "pelican riding a bicycle" benchmark

For anyone building a product on top of an LLM, this is an uncomfortable reminder: any publicly exposed agent leaks clues about the model running behind it, even without documentation or a changelog revealing it. A seemingly silly test prompt can be technical reconnaissance in disguise.

The cost of turning off the tap

Claire decided to end ChatPRD's free trials entirely. For a traditional SaaS, this would mainly cost conversion: the trial is an acquisition engine in product-led growth products, and cutting it raises the entry barrier for legitimate users.

For a product built on top of a language model API, the math is harsher. Every request consumes tokens paid to the model provider, so fraudulent use isn't just lost opportunity, it's direct operating cost (COGS) leaving the cash register with every prompt processed for a game map studio that never intended to pay for anything. Ending the trial stops being a marketing decision and becomes a margin survival decision.

What changes for those building AI products

The ChatPRD case isn't isolated: any commercial wrapper on top of a language model, from code copilots to a vertical agent like this one, carries the same structural risk. It's worth considering a few points before copying the unlimited trial model:

  • Rate limit by usage, not by time. Limiting "30 days free" without limiting token volume per account leaves the door open for those who automate account creation.
  • Monitor prompt patterns at scale, such as the same test text (the equivalent of the pelican on a bicycle) repeated at high volume from distinct accounts, which suggests automated reconnaissance.
  • Detect out-of-scope usage: if the agent was designed to generate PRDs and is being used to generate game text or generic code, that's a sign of tunneling, not legitimate use.
  • Require light verification at sign-up (a card, even without charging it, or phone validation) instead of just email, which is trivially disposable.
  • Limit and rotate promo codes, since promo code farming was an explicit part of the scheme reported by Claire.
  • Consider usage-based billing from the start, instead of a trial with a fixed cap, so infrastructure cost tracks actual usage even during the free period.

What remains open is the real size of the problem beyond ChatPRD: Simon Willison doesn't provide figures on how much this type of fraud has already cost the AI product market as a whole, only the account of a concrete, verifiable case. But the economic logic (expensive model, trial easy to automate, agent too generic to distinguish legitimate use from abuse) suggests ChatPRD is hardly an exception.

Translated from the Brazilian Portuguese original · Read the original

View profile →