More How I AI episodes

How I fight token fraud as a solo founder thumbnail

How I fight token fraud as a solo founder

Published 16 Nov 2026

Duration: 00:39:46

Token fraud exploits AI products via free trials and promo codes, requiring strict spend controls and technical defenses to protect businesses from financial loss.

Episode Description

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...

Overview

The episode examines "token fraud," the systematic abuse of AI product compute by non-customers who exploit free trials, promo codes, and API access to inflate inference costs. The speaker identifies six recurring abuse patterns: free tier farming, promo code resale, bad credit card testing, resale and relay proxies, MCP exploitation, and model probing from international accounts. Specific incidents include an international Roblox and Fortnite map studio running loops of free trials and promo codes, a reseller wrapping the app in a proxy to resell access by the hour, and signup farms using sequential names.

Technical defenses discussed include AI gateways with model allowlists, hourly or live spend monitoring with alerts, Stripe Radar for card fraud, BotID for client verification, firewall and IP blocking, auth provider bot protection, and MCP host restrictions. Operational AI agent loops handle fraud classification of flagged sessions, hourly signup scanning, AI support desk responses, and feedback loops that encode each incident back into rules and playbooks.

Business-level measures carry real costs: eliminating free trials removes a "great source of PLG and enterprise leads," removing promo codes loses partner goodwill, and country-level blocking excludes legitimate users. The speaker notes that "every real suggestion people give me about how I could very easily stop fraud is the exact things that would be bad for our business," framing the core tension between security and growth.

What If

  • What if you treated your free tier as a liability audit rather than a growth channel?

    • Move: Run a 30-day experiment restricting free tier access to business email domains only, capping usage at the cheapest model tier, and tracking conversion-versus-fraud-cost daily.
    • Why Now? The source describes free trials as "the biggest lever" for fraud and the speaker had to shut them off entirely, losing PLG pipeline. A constrained tier may preserve some legitimate conversion.
    • Expected Upside: You could potentially reduce inference costs from "the most expensive models" abuse while retaining a subset of trial-to-paid conversions, avoiding a full trial shutdown.
  • What if you built a fraud classification agent that processes every flagged session before any human review?

    • Move: Implement an LLM-based categorization loop that tags flagged sessions into categories (trial abuse, promo abuse, card fraud, resale, agent farming) and auto-bans or downgrades based on severity tier.
    • Why Now? The source describes this as "super helpful" and notes the speaker's team was spending "80, 90% of our job" on manual fraud triage.
    • Expected Upside: You could potentially reduce manual review time and catch coordinated patterns like the "international Roblox and Fortnite map studio" faster, freeing engineering capacity for product work.
  • What if you treated your MCP and custom integration surface as a primary attack vector rather than a feature?

    • Move: Audit all custom MCP hosts and tunnel endpoints your product supports, restrict to verified corporate domains, and implement per-integration compute budgets with hourly spend alerts.
    • Why Now? The source describes a "monster MCP" that became a "24-7 worker" and the speaker had to stop allowing "free tunnels and hosts for MCP servers that don't seem like they come from real companies."
    • Expected Upside: You could potentially prevent the most resource-intensive abuse pattern (continuous external tool calling) while still supporting legitimate business data connections from verified organizations.

Takeaway

  • Restrict your AI gateway to an explicit model allowlist and enforce per-account and per-org hard usage caps; the source warns that allowing a generic endpoint lets abusers request "Opus, or I want Fable 5.2 instead of Sonnet that you want to give them."

  • Run spend monitoring hourly or live with automated alerts and failovers; the speaker describes waking to "$200,000 had been attempted purchases" and says "I just didn't realize how quickly something could be really exploited and cost would pile up."

  • Implement BotID or equivalent client verification on every chat route to prove requests originate from your app rather than a relay; the source notes this "proves that they're using your app" even when valid credentials are present.

  • Encode every confirmed fraud incident back into your rules, agent playbooks, and account logic so the system learns; the source frames this as "fool me twice, shame on me" and "build it into your rules, build it into your odds, build it to your agents."

  • Restrict MCP and tunnel integrations to verified corporate domains and set per-integration compute budgets; the source describes having to stop allowing "free tunnels and hosts for MCP servers that don't seem like they come from real companies, real concerns."

Recent Episodes of How I AI

More How I AI episodes