Dev.to Security 🔐 Cybersecurity 👁 0 📖 7 min read

If You Sell AI, You Sell Compute: Six Fraud Patterns Every Metered Product Should Check For

In November 2026, on her podcast How I AI, Claire Vo — the solo founder of ChatPRD, a bootstrapped AI platform for product managers — walked through a year of fighting what she calls token fraud. The details are almost c

In November 2026, on her podcast How I AI, Claire Vo — the solo founder of ChatPRD, a bootstrapped AI platform for product managers — walked through a year of fighting what she calls token fraud. The details are almost comic. People used ChatPRD to write product requirements for scamming ChatPRD. Fraudsters routed requests through her product to build Roblox and Fortnite maps. One reseller wrapped the app in a proxy and sold access by the hour. One connected tool — a custom MCP server — ran as a 24/7 compute worker that declared, in its own words, "I'm never sleeping."

The bill for all of it landed on her: tens of thousands of dollars over roughly the past year, and by her own description, almost all of her engineering energy.

If your small business sells anything metered — credits, seats, API access, a free trial on a product with per-use costs — this is not a SaaS horror story from someone else's industry. It's a threat model you already have. Her operating principle is the shortest version of it: "If you sell AI, you sell compute." A seat-based subscription doesn't change the fact that every expensive model call costs real money. Every free trial is a coupon for that compute. So the question that matters isn't "can someone break into our app?" It's: can someone use access we intentionally granted to run up costs our pricing can't support?

Here's the six-pattern checklist from her account, in plain English, and the control stack she layered on top — both worth checking against your own product before the fraud finds it.

The six ways fraudsters farm a metered product

Vo describes six recurring patterns. Run your own product against each:

  1. Trial farming. Repeatedly creating accounts to consume free access — often automated, often deliberately targeting your most expensive models. If a trial touches a per-use cost, a farm is a cost center with a signup form.
  2. Promo-code abuse. Stealing, reselling, or mass-redeeming your partner and marketing offers — sometimes to power someone else's AI business. A promo code is a discount on your margin; at resale scale, it's inventory.
  3. Card abuse. Bad cards, prepaid cards, chargebacks, and card testing. In ChatPRD's case, one morning of card testing meant roughly $200,000 in attempted purchases — all failed, all involving scam or stolen cards. That figure is attempted volume, not a loss, but it shows what the payment layer attracts.
  4. Relays and resellers. Someone authenticates into your product, then routes a different product's requests through it and returns your answers to their own customers. A valid login proves nothing about who the request is really for.
  5. Custom-MCP abuse. If you let customers connect tools or data sources, a connected tool can turn your product into a persistent worker — or bypass the message-level controls you thought were the boundary. That's where the "I'm never sleeping" worker came from.
  6. Model probing. Repeated questions about model identity and training cutoffs, alongside unrelated generation tasks. Vo is honest that she never established what this was — distillation, farming, something else. Treat strange activity as a signal to investigate, not proof of a particular attack.

Notice what all six share: none of them involves breaking in. Every pattern uses access you granted on purpose, for growth. And sophisticated actors, Vo says, combine techniques specifically to stay below each individual threshold you set.

The control stack, in the order it pays back

Business controls first, because they're the ones that change the economics:

  • Charge upfront, and skip grace periods on failed payments. A metered product with a "we'll retry later" window is a free loan program for fraud.
  • Set hard dollar limits per account and per organization — not just request limits. A request cap misses the difference between a cheap model and an expensive one, and between a small workload and a large one. Count dollars, not requests.
  • Put an explicit model allowlist in your application code. Permission to call a provider's endpoint should not become permission to choose any model that endpoint supports — otherwise a session priced for the cheap lane becomes a bill in the expensive one. When ChatPRD suspects fraudulent traffic, it downgrades it to a cheaper model rather than banning on suspicion alone.
  • Review spend hourly, or live. Vo's own admitted mistake was underestimating how fast exploitation accumulates. Cap, alert, and automate a response at the speed the bill can grow.
  • Give your terms of service teeth for suspected abuse, and keep feature flags ready so a vulnerable access path can be disabled quickly.
  • Inspect your existing auth configuration before buying anything new. Her practical warning: useful bot and fraud controls are often disabled by default in the auth setup you already pay for. Ask the provider for help before assuming another purchase is necessary.

For connected tools, constrain the hosts you support — Vo now rejects free tunnels and hosts that don't appear to belong to real companies. For automated traffic, roll out bot detection in the sequence she recommends: observe first, understand your user profiles, tune the control, and only then block. Enforcement you haven't observed is enforcement your legitimate customers will experience first.

One boundary that's easy to miss: your support desk is an attack surface. Fraudsters appeal bans with polished, AI-written explanations, and a persuasive enough message shouldn't be able to reverse enforcement, lift a ban, or mint a fresh promo code. Give support the context to explain and route a decision — not the authority to overturn one. And keep a real appeal path: ChatPRD's enforcement caught at least one false positive in a large ban sweep, and the correction mattered as much as the automation. Enforcement will make mistakes; legitimate customers need a way back in.

The decision that actually mattered: killing the free trial

The most consequential control in the whole account isn't technical. Vo turned off free trials for good — sacrificing a source of product-led growth and enterprise leads — because she couldn't make the economics safe. Requiring a credit card hadn't fixed it: fraudsters cycled cards, used bad ones, and generated chargebacks. Even paid trials weren't sufficient. She's abandoning promo codes too, after watching partner offers become resale inventory.

That's the honest cost math every metered business should run: a signup channel that costs more than it converts isn't a channel — it's a leak. No founder kills a growth funnel for fun. The signal to run the numbers is when fraud defense starts eating your working life — Vo says it has felt like 80–90% of the job.

Geographic blocking is the bluntest version of the same tradeoff. She reports that blocking entire countries significantly reduced fraud, and she says the quiet part too: legitimate users in those countries lose access. That's a business decision to make in advance — how much exposure you can absorb, which signals justify a restriction, and how a wrongly-blocked customer appeals. Don't copy the block; copy the discipline of pricing your own generosity before someone else prices it for you.

Honest boundaries

  • This is one founder's account, not a survey. It's corroborated — her post drew a public response from Atlassian describing the same patterns, noting that every LLM call costs real money and that for small businesses this can be "absolutely crippling" — but your fraud economics depend on your pricing, your market, and your exposure.
  • The classifier details weren't published. The architecture is the copyable part — metadata triggers a narrow investigation, classification informs a scaled response, uncertain cases go to a human — not the thresholds, prompts, or false-positive rates. Don't mistake a model's category label for verified evidence.
  • Agent-assisted defense is another permissions problem. Vo runs agent loops to classify fraud, review signups hourly, and brief the support desk — and each one gets bounded authority. An agent that can ban accounts is an agent that can be talked into banning the wrong one.
  • This is seller-side advice. If you're on the buyer side of a runaway AI bill, the failure mode is different — and so are the controls.

Before any of this: know what you're exposing

Every control above assumes you can answer one question: which parts of your business pay per-use compute bills? The trial on the AI feature inside your product. The promo code on the metered tier. The connected tool a customer installed last quarter. Most owners can't list them — and a defense can't start before the inventory does.

That mapping is what SMB Scale Up's AI Clarity Sprint delivers: two weeks, one fixed fee ($1,495 CAD), a complete inventory of every AI touchpoint in your business, an own-vs-use analysis of what you pay for versus what gets used, three quick wins ranked by ROI, and one first-project plan with an honest build / buy / don't-automate verdict. If you sell anything metered, start by knowing what compute you're already paying for. Details and booking: https://smbscaleup.com/ai-clarity-sprint.html

📰 Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.