←Back/How I AI
How I AI

Token Fraud Is Eating AI Margins. Here’s How Claire Vo Is Fighting It.

After token fraud cost ChatPRD tens of thousands of dollars, Claire Vo rebuilt its defenses across trials, payments, model access, MCPs, and AI support. Here’s what worked—and what legitimate customers lost along the way.

Claire Vo's profile picture

Claire Vo

November 16, 2026·9 min read

Watch or listen to the full episode

Episode chapters

  1. 0:00What token fraud costs a bootstrapped AI businessClaire explains ChatPRD’s seat-based economics and why inference abuse creates direct financial exposure.
  2. 5:24Six abuse patterns, from trial farms to custom MCPsThe observed attacks include promo resale, proxy businesses, card abuse, persistent MCP workers, and model probing.
  3. 13:44Closing trials and tightening business policiesClaire weighs upfront billing, hard usage limits, promo restrictions, geographic blocks, and the legitimate business they can cost.
  4. 23:13Gateways, payment rules, client verification, and MCP controlsThe technical defenses span Vercel AI Gateway, Stripe Radar, BotID, firewall rules, authentication controls, and connected-tool restrictions.
  5. 30:48AI fraud classification and operational review loopsClaire describes threshold-triggered classification, hourly reviews, support guardrails, and feeding new abuse patterns back into enforcement.
  6. 35:13Count dollars, verify clients, and plan for mistakesThe closing lessons focus on compute economics, expensive-model demand, false positives, and the human cost of fraud defense.
Episode outline

People used ChatPRD to write product requirements for scamming ChatPRD. Others routed requests through it to build Roblox and Fortnite maps, or wrapped the app in a proxy and resold access by the hour. One custom MCP became a round-the-clock worker declaring, “I’m never sleeping.”

The absurdity is easier to appreciate when you aren’t paying the inference bill.

In this solo episode of How I AI, Claire Vo describes token fraud that has cost her bootstrapped business tens of thousands of dollars over roughly the past year. Her most consequential response wasn’t a clever detection model: she turned off free trials, sacrificing a source of product-led growth and enterprise leads because she couldn’t make the economics safe.

Her operating principle is blunt: “If you sell AI, you sell compute.” A seat-based subscription doesn’t change the fact that every expensive model call creates a real cost. The defense has to reach beyond login security into billing, model selection, tool access, and support permissions.

The account can be valid while the business model is being abused

ChatPRD charges by seat, with pricing built around average usage, costs, and model routing. That works when customers use the product within its intended bounds. It becomes fragile when someone treats a trial or promotion as a subsidized inference endpoint.

Claire describes six recurring patterns:

  • Trial farming: repeatedly creating accounts to consume free access, often through automation and often targeting expensive models.
  • Promo-code abuse: stealing, reselling, or mass-redeeming partner offers, including to power other AI businesses.
  • Card abuse: bad cards, prepaid cards, chargebacks, and payment-related attacks.
  • Relays and resellers: authenticating in ChatPRD, sending another product’s requests through it, and returning the answers to that product’s customers.
  • Custom-MCP abuse: using connected tools to turn the product into a persistent worker or bypass message-level controls.
  • Model probing: repeatedly asking about model identity and training cutoffs, alongside unrelated generation tasks.

That last category deserves the uncertainty Claire gives it. She doesn’t establish whether the probing was distillation, farming, or something else. Strange activity is a signal to investigate, not proof of a particular attack.

The distinction matters because legitimate usage can look unusual, too. ChatPRD supports coding for prototypes and custom MCP connections for business data. Those useful capabilities also created room for abuse. Claire says sophisticated actors combined techniques to stay below individual thresholds.

She also distinguishes commercial resellers from people seeking AI access where its price is high relative to local incomes. She finds the latter more sympathetic. Neither is something a solo, bootstrapped founder can indefinitely subsidize.

The useful threat-model question is therefore broader than “Can someone break into our app?” It is: Can someone use access we intentionally granted to run up costs our pricing cannot support?

Closing the free door worked—and damaged the growth funnel

Claire’s strongest business judgment is that trials stopped being worth it for ChatPRD. Requiring a credit card hadn’t solved the problem: people cycled cards, used bad ones, and generated chargebacks. Even paid trials, she says, weren’t a sufficient defense.

She is also abandoning promo codes after seeing partner offers become resale inventory. These weren’t disposable marketing experiments. They were ways to put the product into the hands of people she wanted to support and potentially convert.

Her broader controls include charging upfront, avoiding grace periods on failed payments, setting hard limits per account and organization, and using feature flags to disable vulnerable access paths quickly. Terms of Service need to support action against abuse and suspected abuse. Where trials remain available, business-domain restrictions are another possible filter.

Some of these recommendations overlap with Gareth’s post, which Claire credits after hearing that Atlassian was dealing with similar problems. Her own public account of the fraud brought responses from other founders and technical teams experiencing it too.

Geographic restrictions are the bluntest intervention. Claire reports that blocking entire countries through the firewall significantly reduced fraud. She also explicitly acknowledges the cost: legitimate product leaders and engineers in those countries lose access.

That is a reported business tradeoff, not a universal recommendation to copy her country restrictions. An adaptable approach is to decide in advance how much exposure your business can absorb, which signals justify restrictions, and how legitimate users can appeal. Claire’s experience makes the cost of generosity concrete; it doesn’t make every broad block precise.

Control dollars and models before counting messages

The technical stack begins with an AI gateway. Claire uses Vercel AI Gateway for spend controls and control over available models.

The important implementation detail is an explicit model allowlist in application code. Permission to call a provider endpoint should not become permission to choose any model that endpoint supports. Otherwise, an attacker may turn an interaction priced for a cheaper model into a much more expensive one.

Next comes monitoring at the speed the bill can grow. Claire says her mistake was underestimating how quickly exploitation could accumulate costs. Her recommendation is to review spend at least hourly, or live, with caps, alerts, failovers, and automated responses aligned with the business’s economics.

Count dollars, not requests. A request limit alone misses differences in model cost and workload size. Claire also describes downgrading suspected fraudulent traffic to a less expensive model rather than treating every case as an immediate ban.

Payments need their own controls. Claire uses Stripe Radar, initially because of card testing: she woke up to roughly $200,000 in attempted purchases, all failed requests involving scam or stolen cards. That figure was attempted payment volume, not revenue or a reported loss.

Her practical Radar workflow combines an agent with a human review surface:

  1. Use Codex to help implement complex fraud rules.
  2. Review the rules in the Stripe dashboard, where she finds them easier to understand.
  3. Examine Radar’s estimates of what would be blocked and what legitimate payments would pass before enabling them.

She also added card fingerprinting to detect the same card appearing across multiple trials, using a combination of application code and Radar.

These are descriptions of her operating setup, rather than a live implementation tutorial with published commands. The reusable pattern is specific nonetheless: let the coding agent handle rule construction, then review the consequences in the payment system before enforcement. No exact Radar rule or Codex prompt is supplied in the episode.

A valid login does not prove a request came through your product

One reseller was authenticated in a browser while routing requests from its own AI product through ChatPRD. Authentication alone couldn’t answer the relevant question: was this a customer using the app, or someone treating it as infrastructure?

Claire implemented Vercel BotID to assess automated traffic on the chat route. Bot detection is one signal: it does not prove a request is authorized or replace authentication and authorization.

Her rollout advice is to observe traffic, understand user profiles, tune the control, and then block. That sequence is worth copying because it gives you a chance to see what enforcement will affect before making it a customer problem.

Around that layer, she uses Vercel Firewall for country and VPN blocking, IP blocks, and rate limits. She also worked with Clerk to tune authentication protections. Her practical warning: useful bot and fraud controls may be disabled in your existing auth configuration. Inspect those settings and ask the provider for help before assuming another purchase is necessary.

Custom MCP access needed a separate boundary. ChatPRD allowed customers to connect business data through custom MCPs; attackers used that flexibility for external tool calling and compute consumption. Claire now constrains supported hosts, including rejecting free tunnels and hosts that don’t appear to belong to real companies.

The generalizable lesson is to review connected-tool permissions as part of the cost and abuse model. A login check doesn’t constrain what a connected tool can induce your agent to do. In Claire’s case, a feature meant to make customer context accessible also became a route to persistent work at her expense.

Use agents to investigate abuse without giving them the keys back

Claire describes four AI-assisted operational loops. They reduce the manual burden, but they also create another permissions problem to manage.

The first is fraud classification. When chat metadata crosses an abuse threshold, ChatPRD runs what Claire calls a “ZDR JEV loop” to classify the activity. Categories include real customers, trial abuse, promo abuse, card fraud, resale, and agent farming. Depending on the category and tier, the system can ban the account or silently downgrade it to a model such as GLM.

The episode does not provide the classifier prompt, thresholds, evaluation results, or false-positive rate. Copy the architecture cautiously: metadata triggers a narrower investigation, classification informs a response, and uncertain cases need a review path. Don’t mistake a model’s category label for independently verified evidence.

The second loop is an hourly agent reviewing recent signups and unusually out-of-bounds chat threads. High-confidence abuse triggers an automatic ban. Less certain cases go to Claire for feedback or to the AI support desk.

The third loop equips support with fraud lists, allowlists, playbooks, and response guidance. This is where the boundary becomes especially important. People appeal bans, sometimes with polished AI-written explanations. Claire warns that support itself can become an attack surface if a persuasive message can get an account reinstated or produce a fresh promo code.

Support needs enough context to explain and route a decision, without unrestricted authority to reverse enforcement, change bans, or grant access. Claire reports one false positive in a large ban sweep, which the team fixed. That correction matters as much as the automation: enforcement will make mistakes, and legitimate customers need a way back.

The fourth loop feeds newly discovered patterns into application rules, operational checks, and agent instructions. An incident should improve the next detection cycle rather than remain a one-off cleanup.

There is no clean victory claimed here. Claire says fraud defense has felt like 80–90% of the job, and that many effective remedies hurt the business she actually wants to build. Her tools helped implement and operate defenses; they did not remove the need to decide whom to block, how much risk to tolerate, or when customer friction is justified.

The verdict: if your product pays model bills, copy the model allowlist, dollar-based caps, request-level client checks, payment review, and bounded support permissions first. The agent loops are especially useful for small teams that cannot manually inspect every suspicious signup. Closing trials or restricting entire markets requires a separate business judgment. AI can help identify the pattern and implement the rule; a human still owns the cost of getting that rule wrong.

Thanks for supporting How I AI

Sponsors
WorkOS

Drop-in APIs for enterprise features, including authentication, access controls, and audit logs, to help applications become enterprise-ready.

Auros

Auros connects teams with verified human expertise for research and AI.

Build your next product with ChatPRD

Turn an idea into a PRD, user stories, and a plan.

Try ChatPRD free

Start shipping better products.

Join 250,000+ product managers who use ChatPRD to write better docs, align teams faster, and build products users love.

Free to start · 14-day Pro trial · SOC 2 Type II