Free trial abuse in AI apps: credit farming and how to slow it

If your AI app gives free credits, assume someone will automate signing up for them. Slow them with cheap friction at signup, a small and separate free budget, limits on expensive features, and monitoring that spots clusters of new accounts behaving alike. No single control stops a determined farmer.

What credit farming looks like

An attacker creates many accounts, each with its own free allowance, and uses the allowance for their own purposes: running their own product on top of yours, reselling access, or generating content at scale. Your model bill pays for it.

OWASP's API list describes the general shape in its sensitive business flows and resource consumption pages: an abuse of a legitimate flow at volume, where each request on its own looks normal. The resource consumption page also notes that third-party services that charge per request raise the stakes, since attackers can drive up costs.

Cap the damage first

Before adding any detection, make the worst case small.

1. Give free users a modest, fixed budget in tokens or cost, not unlimited requests.
2. Fund it from a separate key with its own spend limit. See separating keys per environment and feature.
3. Restrict the expensive features on free plans: longer inputs, larger models, file uploads, tool use, image generation.
4. Cap output length and concurrent requests per account.

If the whole free pool has a hard monthly ceiling, an attack becomes a bounded cost and not an incident.

Add friction at signup

• Verify email, and treat disposable email domains with suspicion.
• Require a verified phone number or a payment method on file for higher free tiers. Be aware that this costs you some signups.
• Use a bot challenge on signup and on the first model request, not on every request.
• Delay full free credits until a short, real action, such as completing a setup step.

None of these is perfect. Each raises the cost for the attacker, and the aim is to make farming less profitable than the attacker's alternatives.

Look for patterns across accounts

Individual accounts look fine. The group does not. Signals worth watching:

• Many signups from the same network range or device fingerprint in a short window
• Accounts that use their full free budget within minutes of creation
• Near-identical prompts across accounts
• Accounts that never use the product's other features
• Traffic patterns that look scripted, such as even timing

Flag these for review and slow them, rather than banning outright. A review queue lets you correct false positives.

Handling the account you catch

Suspend the free allowance, keep the account for evidence, and tell the user how to appeal. Have a written acceptable use policy that names automated signups and resale. See writing an acceptable use policy for an AI product.

Do not forget the real users

Heavy friction at signup loses honest customers. Measure signup completion after each change. Often the cheapest fix is the budget cap, which costs honest users nothing.

Quick check

Sources

• OWASP API4:2023 Unrestricted Resource Consumption

Last reviewed: 19 October 2026.