Token fraud is now a business-ending cost for small AI products, and ChatPRD founder Claire Vo says she ended free trials to stop it. Vo runs ChatPRD, an AI app that helps product managers turn an idea into a spec. In a YouTube episode of How I AI she describes a flood of abuse from people farming free trials, some for free AI usage and some to resell it or run other businesses on top of her agent. Simon Willison flagged the video on November 20, 2026. ChatPRD's own write-up of the episode says the fraud cost the bootstrapped company "tens of thousands of dollars" over about a year.
This post covers the six patterns she lists, what Stripe says about the wider trend, and the defenses an AI app can copy. All dollar figures here come from Vo's own account and are not independently verified.
Token fraud at a glance #
| Question | Answer |
|---|---|
| What is it? | Getting AI compute without paying, often to resell it |
| Who reported it? | Claire Vo, founder of ChatPRD, on the How I AI show and in ChatPRD's write-up |
| Cost to ChatPRD | "Tens of thousands of dollars" over roughly a year, per Vo; no exact figure given |
| Her main response | Ended free trials and promo codes; added caps, allowlists and monitoring |
| Wider data | Stripe executives say trial abuse more than doubled in six months (Fortune, May 2026) |
| Open question | The true size across the industry; the numbers come from vendors and one founder |
Agents that act on your behalf need rules. See how explainx helps secure your agents →
What does token fraud look like in practice? #
ChatPRD's write-up lists six patterns. Each one turns a normal product feature into free compute.
- Trial farming. People create repeat accounts to consume free access, often with automation, and often aimed at the most expensive models.
- Promo-code abuse. Partner offers get stolen, resold or mass redeemed, sometimes to power other AI businesses.
- Card abuse. Bad or prepaid cards, cycled cards, chargebacks and card-testing attacks. The write-up mentions about $200,000 in attempted purchases that all failed, and says this was attempted payment volume, not revenue or a reported loss.
- Relays and resellers. Another product routes its users' requests through ChatPRD and returns the answers, so ChatPRD pays for someone else's customers.
- Custom MCP abuse. Users connected tools through the Model Context Protocol to create a persistent worker or to bypass message-level controls. One was described as "I'm never sleeping."
- Model probing. Repeated questions about model identity and training cutoffs, mixed with unrelated generation tasks. Vo does not claim to know the intent: it could be distillation, farming or something else.
One example from the video, as summarized by Willison: an abuser said they were funding an international Roblox and Fortnite map studio using free trials and promo codes to build maps through her agent. Willison adds that model families have distinctive habits, such as how they draw an SVG of a pelican riding a bicycle, and that these can help identify the model behind an app.
For readers new to the term, MCP is explained in our guide.
What does Stripe say about token fraud? #
Vo is not alone. In a May 2026 Fortune report on Stripe, Stripe CEO Patrick Collison said on the TBPN podcast that token thieves make up about one in six new signups at AI firms. Emily Sands, Stripe's head of data and AI, called token theft "the most under-discussed topic in AI" and said free-trial abuse had more than doubled over six months. She compared it to "dine and dash": fraudsters open accounts across companies and burn tokens at machine speed, often with agents.
Two caveats from the same report. A startup whose customer-acquisition cost rose to $500 because of trial abuse was a single example, not an industry figure. And a Stripe client's conversion rate rising from 1 in 25 to 1 in 3 after blocking fraudulent accounts was also specific to that client. Stripe is adapting Radar, its fraud detector, to flag token fraud. These are Stripe's statements and the report notes both examples are case-specific.
What defenses did ChatPRD actually use? #
According to ChatPRD's write-up, these are the changes Vo made.
A coin jar with a clamped lid and a weekly budget tag, representing a spend cap against token fraud
Business rules
- Ended free trials. Requiring a card and running paid trials did not solve the problem.
- Ended promo codes, because partner offers had become resale inventory.
- Charge upfront, no grace periods on failed payments, hard per-account and per-organization limits, and feature flags to switch off vulnerable paths quickly.
Technical controls
- Model allowlist in application code , so access to a provider does not mean access to any model.
- Dollar-based spend monitoring at least hourly, with caps, alerts and failovers. The advice is to count dollars, not requests. Suspected fraud can be downgraded to a cheaper model instead of banned.
- Stripe Radar rules against card testing, plus card fingerprinting to catch one card used across multiple trials. Vo used Codex to draft rules, then reviewed them in the Stripe dashboard and tested them against Radar's block estimates before enabling.
- Vercel BotID and Firewall , rolled out in stages: observe, understand user profiles, tune, then block. Country and VPN blocks, IP blocks and rate limits are part of it.
- MCP host restrictions that reject free tunnels and hosts that do not look like real companies.
Country blocking significantly reduced fraud, per Vo, but legitimate users there lost access. The write-up frames it as her trade-off and not a recommendation.
How is AI used to fight the fraud? #
Vo also uses AI in the defense. When chat metadata crosses a threshold, a classification loop sorts activity into groups such as real customer, trial abuse or card fraud, and it can ban or downgrade. An agent reviews recent signups and odd threads every hour. High-confidence abuse is auto-banned. Uncertain cases go to Vo or the support desk.
Support staff get fraud lists and playbooks but limited authority, because persuasive appeals could otherwise reinstate accounts or produce promo codes. Vo reports one false positive in a large ban sweep, which was fixed. The write-up gives no prompts, thresholds or false-positive rates, so this is a pattern to copy, not a benchmark.
Where does this connect to other AI security stories? #
Stolen and abused model access is a recurring theme. Amazon's decoy APIs recorded stolen-key campaigns by agentic tools, and Censys counted 1,465 exposed ARTEX consoles. Those cases target keys and APIs. Vo's case targets the signup flow, but the goal is the same: someone else's compute at no cost. For the demand side, OpenRouter token volume after Claude 5.5 shows how quickly usage can grow, which also raises the stakes of an unmonitored free tier.
What should an AI app builder do now? #
Vo's recommended first steps, from the write-up: add a model allowlist, dollar-based caps, client checks, payment review and bounded support permissions. Closing trials or blocking markets is a separate business decision, because both cost growth.
A practical checklist:
- Cap spend per account and per organization in dollars , not messages. A single agent loop can burn a month of quota in an hour.
- Allowlist models in code. Do not let a user-supplied model name reach the provider.
- Alert hourly on spend , and downgrade suspected accounts to a cheaper model before banning.
- Treat promo codes as currency. Track redemptions by source and expiry.
- Restrict what connected tools can do. Give custom MCP access its own boundary and reject throwaway hosts.
- Limit support authority. Appeals should not be able to restore access or create credits.
- Review bans. Log every automated ban so you can find false positives.
If your app lets AI agents act with real credentials and spending power, you also need to control what those agents do. AgentBeam, the agent security platform from the explainx.ai team, is built to stop AI agents before they take dangerous actions.
What does this mean for users and founders? #
For users, the visible effect is fewer free tiers and shorter trials. For founders, the lesson is that "if you sell AI, you sell compute," in Vo's words, and compute attracts the same kind of abuse that cloud credits and SMS verification attracted before. The fraud figures in this story come from one founder and from Stripe, a payments vendor with a product to sell. Take the direction as credible and the exact numbers as unconfirmed. The trade-off is hard for a small team. Free trials help growth, but each trial account can carry a real model bill. Vo says ending trials cost her growth and enterprise leads, yet it stopped the losses. A larger company may absorb that cost, while a bootstrapped one often cannot. Builders should decide early how much free compute they can afford to lose, and set caps and monitoring before they launch a trial, not after the first large bill arrives.
Related reading #
Figures are as reported by Claire Vo, ChatPRD, Stripe and Fortune as of November 20, 2026, and are not independently verified.