The AI API abuse playbook: how farmers drain free tokens and how to stop them

Why AI products are a special target

AI products give away something that has a direct resale value: model access. A free tier of a coding assistant, image generator or chat API is useful to the person who signs up, but it is even more useful to someone who can pool hundreds of free tiers into one stream of free inference.

That is why abuse on AI products looks more organised than abuse on classic SaaS. There are scripts that automate signup, proxies that rotate IP addresses, and marketplaces where farmed accounts or API keys are resold.

Stage 1: Account creation at scale

The farmer needs many accounts. They use email tricks, disposable inboxes, residential or data-center proxies, and headless browsers driven by automation tools. Signup flows that finish in a few seconds and look identical every time are the result.

Defence: invisible proof-of-work, headless browser detection, signup timing checks, disposable domain lists, mailbox normalisation and device linking. None of these needs a visible CAPTCHA.

Stage 2: Draining the allowance

Once an account exists, the farmer burns through the free allowance as fast as possible, often via the API rather than the interface. Usage is bursty, highly repetitive and starts immediately after signup.

Defence: stagger the allowance, cap free usage per hour, and require a second signal before unlocking the full amount. A real person exploring your product rarely uses the entire free tier in the first ten minutes.

Stage 3: Pooling and reselling

Advanced operations pool many free accounts behind a proxy that spreads requests across them, effectively building an unofficial free API on top of yours. Others sell the accounts or keys outright.

Defence: link accounts by device, network and mailbox; watch for many accounts calling the API from the same infrastructure; and revoke free credits across an entire linked group at once. Unfarm's Farming rings view shows these groups with the shared signals that connect them.

Stage 4: Adapting to your rules

Farmers test your defences. When one trick stops working, they switch to another: from plus tags to numbered series, from data-center IPs to residential proxies, from one device to a small set of devices.

Defence: rely on combinations of signals rather than single rules, and review outcomes. Mark accounts as farming or legitimate, and look at which reasons actually predicted abuse. Rules that never fire on confirmed farmers, or fire often on legitimate users, should be adjusted.

A layered defence that respects real users

Layer 1 at signup: invisible checks, email normalisation, disposable domains, device and network linking. Layer 2 at credit grant: small instant allowance, full allowance after a real-person signal. Layer 3 during usage: rate caps on free plans and alerts on unusual bursts. Layer 4 after the fact: group linked accounts, review, and act on the whole ring.

At every layer, show the reasons. A developer who is held for review and sees that their device already has two free accounts usually understands. A developer who is blocked with no explanation usually churns.

Where to start this week

First, measure: run your last few thousand signup emails through a grouping check to see how many are repeat accounts. Second, stagger your free allowance so the full amount is not available instantly. Third, add device and email checks at signup and hold, rather than block, risky accounts. These three steps usually remove most of the cost of farming without any visible change for honest users.