Skip to content

By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

How I AIHow I AI

Token fraud is here and it’s BAD. Here’s what you need to know.

Today I’m walking you through a problem that has cost me tens of thousands of dollars over the past year and eaten almost all of my engineering energy: token fraud.

Claire Vohost
Nov 16, 202639mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

Great conversations run long. Their best ideas don't have to.

uListen is free on iPhone, iPad, Android, Chrome, and Claude.

See how

CHAPTERS

Jump to publisher timestamps

Chapter summaries

  1. Token fraud hits ChatPRD: what it is and why it’s so costly

    Claire Vo defines token fraud as non-customers (or abusive users) consuming paid LLM API tokens through her product in ways that destroy unit economics. She explains how seats-based pricing can be exploited when usage is unbounded, and shares that it has cost her tens of thousands of dollars as a bootstrapped solo founder.

    • •Token fraud is abuse of an AI product’s LLM API usage that drives up inference costs
    • •Seats-based pricing can be undermined when token consumption isn’t effectively constrained
    • •The fraud has diverted most engineering time toward blocking abuse
    • •The financial impact is especially painful for a bootstrapped solo founder
  2. How AI products get abused: free tier and trial farming

    Claire begins a breakdown of common abuse patterns she’s seen, starting with free tier and trial farming. She explains how generous trials attract automation and heavy use of expensive models for work unrelated to the product’s intended purpose.

    • •Fraudsters repeatedly sign up for free/trial access to extract expensive model usage
    • •Automation turns “PLG generosity” into a predictable attack surface
    • •Turning off free trials can reduce abuse but harms legitimate pipeline and conversion
    • •Abuse often involves using the most expensive models rather than product-aligned work
  3. Promo code abuse and payment chaos: reselling, stealing, and chargebacks

    Claire describes promo code abuse as a major attack vector, including reselling and mass use to power other AI businesses. She also covers payment-related abuse like bad cards, prepaid cards, and chargebacks—painful but more familiar to handle than token fraud itself.

    • •Promo codes can be stolen, resold, and used at scale to extract AI compute
    • •Abusers may spin up other AI businesses using discounted access
    • •Bad card behavior includes prepaid cards, chargebacks, and other payment fraud patterns
    • •Some payment fraud is easier to control than token abuse, but still costly and time-consuming
  4. Resell/relay schemes, reverse proxies, and MCP-based bypasses

    Claire shares unusual and creative ways abusers resell access, including routing other products’ requests through ChatPRD via browsers and proxies. She also recounts abuse using custom MCP setups to run constant worker-like prompting and bypass message-level controls.

    • •Abusers can relay requests through an authenticated session and return results to their own customers
    • •Reverse proxies and relays can be inefficient but still damaging at scale
    • •Custom MCP installations can become 24/7 workers that continuously consume compute
    • •Message-level controls can be bypassed through unexpected integration paths
  5. Probing and “weird” usage signals: separating fraud from legitimate privacy-first use

    Claire notes model probing behavior (questions about model identity and cutoff) and other odd prompt patterns, which are hard to interpret without inspecting user data. She explains the challenge of detecting fraud when the product doesn’t retain or review customer content, and when abusers blend techniques to stay under thresholds.

    • •Model probing prompts can indicate abuse but aren’t always definitive
    • •Privacy posture makes it harder to inspect content to confirm intent
    • •Only extreme examples provide enough signal to investigate confidently
    • •Sophisticated abusers combine multiple techniques to stay below detection thresholds
  6. Real-world stories: map studios, hourly resellers, signup farms, and “Temu ChatGPT” usage

    Claire gives concrete examples of token abuse, including loops of trials and promo codes used to build Roblox/Fortnite maps, and resellers wrapping the app in a proxy to sell access by the hour. She also describes signup farms and general trial misuse where people treat the product as a cheap general-purpose ChatGPT.

    • •Trial/promo loops used to fund external commercial work (e.g., map creation)
    • •Proxy-based resellers selling access to ChatPRD by the hour
    • •Signup farms generate many near-identical accounts in sequence
    • •General-purpose misuse turns a niche tool into a subsidized chatbot
  7. Who’s doing it: criminal enterprises vs. more “sympathetic” trial abuse

    Claire distinguishes between the most damaging abusers—commercial or coordinated criminal efforts—and less nefarious misuse driven by affordability gaps. While she’s more sympathetic to the latter, she emphasizes she can’t subsidize it as a business.

    • •The worst abusers are commercial entities coordinating to extract value
    • •Some trial abuse appears driven by local economics where AI costs outstrip income
    • •Sympathy doesn’t change the reality that the business can’t fund that usage
    • •Different motivations still require strong controls to prevent runaway bills
  8. Business controls that feel bad but work: shutting doors, tightening billing, and limiting access

    Drawing on shared experiences (including a list from Gareth at Atlassian), Claire outlines business-side measures that reduce fraud but also reduce growth. She argues free trials may no longer be worth it, promo codes require paranoia, and strict limits and upfront billing are increasingly necessary.

    • •Free trials—even card-backed—can be cycled and abused; she believes trials are effectively “over”
    • •Charge up front with no grace periods on failed payments to reduce abuse windows
    • •Set hard usage limits at account and organization levels even if it hurts UX
    • •Restrict trials to business domains and use feature flags to quickly disable risky surfaces
  9. Hard choices: country-level blocking, terms of service, and reporting cybercrime

    Claire describes blunt but effective measures like firewalling entire countries and rethinking international strategy. She also mentions ensuring terms of service allow banning suspected abuse, filing cybercrime reports, and encouraging founders to discuss token fraud patterns openly.

    • •Country-level firewall blocks can dramatically reduce fraud, but exclude legitimate users
    • •Terms of service should explicitly allow bans for abuse and suspected abuse
    • •She filed cybercrime reports after coordinated attempts to bypass controls
    • •Public discussion and founder-to-founder sharing helps identify cross-product proxy patterns
  10. Technical defenses: gateways, model allowlists, and spend monitoring that matches token economics

    Claire shifts to engineering controls, emphasizing that spend must be monitored aggressively because costs can spike quickly. She recommends an AI gateway to control routing and available models, plus alerts, caps, and automated failovers aligned with the business’s risk tolerance.

    • •Use an AI gateway to control spend and constrain which models can be called
    • •Implement model allowlists to prevent users from selecting more expensive models implicitly
    • •Monitor spend hourly or live; rapid exploitation can accumulate costs fast
    • •Add automated alerts, caps, and failover behaviors when spend deviates from expectations
  11. Payments and traffic controls: Stripe Radar, Bot ID, firewalling, and auth-provider protections

    Claire explains how she uses Stripe Radar for fraud rules and card fingerprinting across trials, and how she leaned on tools to detect bot-like traffic. She also highlights Vercel Bot ID, firewall rules, rate limiting, and turning on protective settings in auth providers like Clerk, plus controls available from WorkOS.

    • •Stripe Radar helps block suspicious payments, countries, and repeated card reuse across trials
    • •Card testing attacks can create massive attempted charges even when they fail
    • •Bot ID can prove requests originate from the real app, not just stolen credentials
    • •Firewall, IP blocking, and rate limiting reduce exposure before requests hit costly inference
    • •Auth providers often have bot/fraud protections that are off by default and can be enabled
  12. MCP and integration monitoring: constraining custom hosts and blocking risky tunnels

    Claire returns to MCP-specific abuse, explaining how custom MCP hosts and tunnels can become a channel for constant tool calling and compute drain. She describes tightening which MCP hosts are supported and disallowing “free tunnels” that don’t look like they come from real companies.

    • •Monitor not only app usage but also user-installed MCPs and their behavior
    • •Custom MCP hosts can trigger heavy external tool calling and sustained compute usage
    • •Limiting supported MCP hosts reduces attack surface but restricts flexibility
    • •Blocking suspicious tunnels helps prevent stealthy always-on compute extraction
  13. Bots that watch the bots: agent loops for classification, review, and support hardening

    Claire outlines operational “agent loops” she uses to fight fraud: classification of suspicious chats, automated review of signups and threads, and support workflows that can respond without becoming an attack vector. She stresses that support systems must not have the power to undo bans or issue access that abusers can exploit.

    • •A decision-model loop categorizes suspicious activity into fraud types and triggers actions
    • •Suspected abuse can lead to bans or silent downgrades to cheaper models
    • •An hourly agent reviews recent signups and out-of-bounds chats to detect patterns early
    • •AI support needs fraud playbooks and guardrails so it can’t be tricked into unbanning or granting access
    • •Fraud signals should feed back into rules and automation to reduce repeat incidents
  14. Final takeaways: treat AI as compute, close zero-dollar doors, and expect false positives

    Claire concludes that selling AI means selling expensive compute, so businesses must track dollars, not just requests. She warns that trials, coupons, and powerful models attract abuse, that AI tools can be used to plan scams, and that policies must be strong even when they create false positives.

    • •If you sell AI, you sell compute—watch the bill in dollars, not request counts
    • •Every zero-dollar entry point (trials/coupons) will be exploited
    • •Expensive models are a primary target; turning them off in some regions reduced abuse
    • •Fraudsters may use the product’s own AI to help design scams against it
    • •Verify the client beyond login, accept that false positives happen, and keep rules simple and firm

Publisher timestamps

  1. 0:00Token fraud and what it cost me
  2. 5:25Six abuse patterns
  3. 11:24Four kinds of abusers
  4. 13:45Business controlsRead this chapter
  5. 21:27Technical defensesRead this chapter
  6. 30:25Bots that watch the bots
  7. 35:30Final takeaways

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.