How I fight token fraud as an AI founder
Claire, founder of ChatPRD, discusses the widespread problem of token fraud—where bad actors abuse free trials, promo codes, and other mechanisms to steal API inference at the expense of AI product companies. She shares six types of abuse she's experienced, specific fraud examples, and details both technical and operational countermeasures she's implemented to protect her bootstrapped business.
Summary
Claire, a solo founder of ChatPRD (an AI platform for product managers), reveals that token fraud—the abuse of AI API calls through free trials, promo codes, and other exploits—has cost her tens of thousands of dollars over the past year and nearly made her reconsider running the business. She discovered this problem affects not just small bootstrapped companies but also major enterprises like Atlassian, OpenAI, and others in the top 10 AI products.
She identifies six primary fraud vectors: (1) free tier/trial farming, where fraudsters sign up repeatedly for free access and use expensive models for non-product work; (2) promo code abuse, including reselling and mass redemption; (3) bad credit card fraud involving chargebacks and prepaid cards; (4) resale and relay schemes where bad actors wrap her product in proxies and resell it; (5) model probing from international accounts attempting to discover model details; and (6) creative misuse like funding Roblox/Fortnite map studios through trial loops.
Specific examples include an international Roblox studio running continuous free trial and promo code loops to generate map code, a reseller who wrapped ChatPRD in a proxy and sold hourly access to customers, and sophisticated fraudsters building monster MCPs (Model Context Protocols) to bypass message-level controls. Some fraudsters were even running reverse proxies simultaneously across multiple top-10 AI products, suggesting coordinated fraud rings.
On the business side, Claire has had to implement tough policies: eliminating free trials entirely (despite their value for product-led growth), discontinuing promo codes, requiring upfront payment with no grace periods, implementing hard usage caps per account, restricting trials to business email domains, and even blocking entire countries via firewall when fraud signals were high. She also filed FBI cybercrime reports for what appeared to be coordinated international fraud enterprises.
Technically, she uses: an AI gateway (Vercel AI) to control which models are available and enforce spend limits; Stripe Radar for card testing and fraud pattern detection; Vercel Firewall for IP blocking and rate limiting; Bot ID from Vercel to verify requests originate from her app; auth provider protections like bot detection from Clerk; and monitoring of custom MCPs that users install. She also uses AI agents (Claude/Opus) to automatically classify fraud signals and categorize abuse types, with automated bans for high-confidence fraud.
Clair emphasizes that this fight has become 80-90% of her operational focus and that every fraud prevention measure has negative business consequences. However, she stresses that treating AI products like compute businesses (monitoring model costs in dollars, not requests) and starting from a "don't trust you" posture that users must earn is essential. She calls on other founders to share fraud patterns and lock arms against coordinated abuse.
Key Insights
- Free trials, once a valuable product-led growth lever, have become economically untenable for AI products because fraudsters farm them more effectively than legitimate users convert, forcing many companies to eliminate trials entirely despite business impact.
- Sophisticated fraudsters are running coordinated reverse proxy schemes simultaneously across multiple top-10 AI products, reselling tunneled API access to their own customers, suggesting organized fraud rings rather than isolated bad actors.
- AI agents using decision models (like Claude Opus as a classifier) can cost-effectively categorize fraud by analyzing chat metadata and abuse signals, then automatically downgrade suspected fraudulent requests to cheaper models or trigger bans.
- Blanket country-level firewall blocking based on fraud signal analysis reduced abuse significantly for ChatPRD, demonstrating that aggressive blunt-instrument approaches are more effective than nuanced, graduated fraud responses for small bootstrapped companies.
- Fraudsters use the target AI product's own capabilities—like ChatPRD's PRD-writing features—to generate detailed plans for how to scam the platform, turning the creator's tools against their own business.
Topics
Transcript
[0:00] If you think AI slop is bad, let me introduce you to my new very unwelcome best friend, token fraud. In addition to being a lovely AI podcast host, I am the solo founder of Chat urity, an AI platform for product managers that I've been running as [music] a labor of love and a beautiful little SaaS product since I think the early GPT 3.5 days. And I have to tell you, the last year I have been spending more and more of my time and almost all [0:32] of our engineering energy focused on blocking one thing, token fraud. We are going to talk today all about it, what it is, how to spot it, how you…
Full transcript available for MurmurCast members
Sign Up to AccessMore from How I AI
Fraudsters wrote PRDs on how to scam my AI product
An AI product founder shares critical lessons about preventing fraud and abuse of AI services, emphasizing that fraudsters actively create detailed plans to exploit AI platforms. Key defenses include monitoring costs instead of request counts, verifying clients thoroughly, and implementing strict policies with no exceptions.
I use Stripe Radar to catch card testing before it costs me
The speaker explains how to use Stripe Radar to prevent card testing fraud by leveraging built-in fraud rules, fingerprint detection, and Spot ID verification. These tools help identify suspicious payment patterns and validate that requests originate from legitimate app users, while allowing tuning to minimize false positives for good users.
Token fraudsters relayed their AI product through ChatPRD
The founder of Chat Purity discusses token fraud, where bad actors authenticate as their service and relay requests from their own AI products through Chat Purity to appear legitimate. Despite existing technical controls, fraudsters employed reverse proxies and relay systems to obscure their activity, creating inefficient and suspicious infrastructure.
He wants AI that keeps students doing the cognitive effort in math
The speaker emphasizes that AI in mathematics education should support student agency and cognitive effort rather than providing direct answers. Effective AI tools should interpret student ideas, encourage deeper thinking, and engage students in productive struggle while keeping them as active decision-makers in their learning process.
He put the family to-dos on a wall display Claude built
The speaker uses Dakboard, a wall display product, to make personal to-do items and information visible to their family. They had Claude build a script to scrape to-dos, upcoming trips, and other data from their database to display on the wall, including sections for immediate tasks, longer-term needs, and recent wins.