Kavra LabPOST /v1/api-keysVerifyNew key from a device seen on 4 trials
- Device linked to 4 trial accounts
- Disposable email domain
- Real browser, human input
Risk64
Industry
SaaS fraud prevention that stops credit farming, not developers
SaaS and AI fraud is abuse of software products through their free tiers, signups and APIs: endless free trials, farmed LLM and GPU credits, stolen or resold API keys, fake signups that inflate metrics and bots scraping generated output. Kavra assesses every signup, key and request, links accounts to one actor and lets real developers straight through.
- Who it hits
- AI apps, LLM APIs, dev tools, B2B and B2C SaaS
- What it costs
- Compute bills, skewed metrics, lost upgrades
- Tools used
- Scripts, headless browsers, proxies, temp emails
- Where to stop it
- Signup, credit grant, key creation, API
What is SaaS and AI fraud?
Most software grows through a generous front door: a free tier, a trial, starter credits, an API key in two clicks. SaaS fraud is abuse of that door. One person opens the tenth trial, a script creates a thousand accounts to pool free credits, a leaked API key ends up behind a paid proxy someone else sells, or a bot harvests the content your product generates.
AI products changed the math. A classic SaaS trial costs little to serve, so an abuser mostly costs you a lost sale. An AI trial runs on GPUs. Every farmed credit is real inference you pay a model provider or a cloud bill for, which turns free-trial and credit abuse from a leak into a direct cost that scales with the attacker's patience.
Threats that hit SaaS and AI products
The same product surface serves developers, buyers and attackers. Here is where each threat lands.
| Threat | Where it strikes | Business impact |
|---|---|---|
| Free-trial and credit farming | Signup, trial start, credit grant | GPU and inference spend with no revenue |
| Repeat trials through multi-accounting | Signup, workspace creation | Paid plans never bought, cohorts polluted |
| Fake account creation at scale | Signup, email and phone verification | Inflated signups and active users, noisy funnels |
| API abuse, key sharing and reselling | API key creation, API traffic | Quota drained, capacity taken from paying users |
| Scraping of generated content | Public share pages, output endpoints | Outputs copied or used to train rival models |
| Account sharing | Login, concurrent sessions | One seat paid, a team served |
| Account takeover and credential stuffing | Login, key and billing settings | Stolen keys, data exposure, surprise bills |
| Card testing on checkout | Upgrade and billing forms | Disputes, fees and processor risk |
| SMS pumping | Phone verification | SMS bills for traffic that never becomes users |
How credit farming works
Farming runs like a small production line. Each step defeats one classic check.
- 01
Script the signup
A headless browser or an HTTP client posing as a browser fills the signup form. Disposable inboxes and plus-address or dot tricks on one mailbox pass email verification.
- 02
Look like a new person each time
Every account gets a fresh browser profile, often from an antidetect browser or a cloud virtual machine, and a different IP from a residential proxy pool or a VPN.
- 03
Claim the credits
The script activates the trial, accepts the starter credits and creates an API key. Some operators also chain referral links between their own accounts, a form of promo abuse, to collect bonus credits on both ends.
- 04
Pool the keys
Hundreds of keys are loaded into a rotating pool behind one proxy endpoint, so each key stays under its own rate limit while the pool serves heavy traffic.
- 05
Use or resell
The pooled capacity runs the operator's own workload or is resold as cheap access to your models. When keys run dry, the line starts again.
Touchpoints to protect
Developers hate friction, and they are your best customers. The goal is not to put a gate at every step, but to assess each step quietly and add a check only when the evidence asks for one.
- Signup and workspace creation: is this a new person, or an actor already holding trial accounts?
- Email and phone verification: is a real inbox or number behind it, or a disposable one, and is the phone step being pumped?
- Trial start and credit grant: should this account get the full credit amount now, a smaller amount, or credits after a card?
- API key creation: is the key being created by a person in a real browser or by the same script that made the last fifty accounts?
- API traffic: does the pattern look like one app, or like a pool of keys fronted by a reseller?
- Login and session: does the login match the account's own devices, or are several people on different continents sharing one seat?
- Upgrade and billing: is the checkout a buyer, or a bot testing stolen cards against your cheapest plan?
Signals that separate abusers from real users
Developers use VPNs, terminals and new laptops. None of that is fraud alone. Contradictions and links between accounts are.
Many trials, one device
New accounts keep appearing from a device already tied to earlier trials, even after the fingerprint is rotated.
Automation posing as a browser
The signup claims to be Chrome on a laptop but behaves like a scripted client: no real rendering, instant typing, identical timing.
Proxy exits in the right country
Each account looks local, but the IPs belong to commercial proxy pools rather than the home or office networks they claim.
Keys that act like a pool
Many keys from unrelated accounts call the API with the same prompts, rhythm and client, from shared infrastructure.
One seat, many people
One login active on several devices in distant places at the same time, far beyond normal travel or a second laptop.
Agents that claim a name
Traffic says it is a known AI agent but comes from outside the operator's published ranges and carries no valid signature.
Why usual trial defenses fall short
Each common control stops one layer, and each adds friction that honest developers feel first.
| Defense | What it stops | What gets through, or what it costs |
|---|---|---|
| Email domain blocklists | Known disposable inboxes | New throwaway domains and alias tricks on real mailboxes |
| Phone verification | Cheap bulk signups | Virtual numbers, SMS pumping and drop-off from real users |
| Card required for trial | Casual repeat trials | Stolen and prepaid cards, which adds payment fraud, and fewer honest signups |
| Per-key rate limits | One heavy key | Key pools that spread load across many accounts |
| IP limits | Signups from one address | Rotating residential and mobile IPs |
| CAPTCHA on signup | Crude scripts | Solving services, plus friction on developer onboarding |
Business and compliance context
For AI products, free usage is a line on the compute bill, so abuse shows up as margin loss, not only as churn. It also skews the numbers you run the company on. Fake signups inflate activation and active-user counts, pull conversion rates down and send growth spend toward channels that bring bots. Keys resold by third parties can also expose you to use cases that break your model provider's policies or your own.
Protection has to respect privacy law as well. Under the GDPR and similar rules, device signals used for fraud prevention need a legal basis and must follow the user's consent choices where they apply. Opaque identifiers, strict tenant isolation and clear retention rules make fraud checks easier to defend to customers, auditors and your own security reviewers.
What good protection looks like for SaaS and AI
The best setups keep the front door wide for real developers and make farming more expensive than paying.
- Assess every signup and every key creation invisibly. No puzzles for the honest majority.
- Link accounts to the actor behind them, so the fifth trial from one device is known as the fifth.
- Grade the response: full credits for clean signups, reduced or delayed credits for mixed evidence, a card or step-up only for the risky few.
- Score API traffic server side, not only in the browser, so key pools and resellers show up by behavior.
- Recognize verified AI agents and crawlers, and decide per path whether to allow, check or block them.
- Watch sessions for account sharing, and compare each login with the account's trusted devices.
- Run in observe-only mode first, and review what you would have stopped before you change credit rules.
Sources
How Kavra helps
How Kavra protects SaaS and AI products
One script on signup and one server call per decision. Kavra analyzes 3,000+ data points and returns an explained verdict your backend can act on.
Trials linked to one actor
Returning devices are recognized across accounts, and fingerprint rotation is kept as one actor with N rotations, not N new users.
Scripts and spoofed browsers caught
Headless browsers, automation frameworks, HTTP clients posing as browsers and antidetect profiles are exposed by contradictions between layers.
Server-side API assessment
A request API scores backend and API traffic, so key pools and resellers are caught where they actually hit you.
Verified AI agents recognized
Known agents and crawlers are verified by signature and published ranges. You choose to allow, check or block each one.
Friction only for the risky few
No CAPTCHA puzzles. Invisible challenges run first when evidence is unclear, and your rules decide the step-up.
Built for developers
A script under 64 KB that never blocks rendering, one API call, webhooks and native iOS and Android SDKs. First results the same day.
FAQ
Frequently asked questions
Something else? Talk to our team.
How do I stop people from creating multiple free trials?
Recognize the person, not the email. Repeat trialists change inboxes, IPs and browser profiles, but their device, network and behavior tend to link back to earlier accounts. Assess each signup for those links and act before credits are granted: allow clean signups, reduce or delay credits for mixed evidence and ask for a card only when the account ties to earlier trials.
What is credit farming in AI apps?
Credit farming is creating many accounts to collect the free credits each one receives, then pooling them to run workloads or resell access. In AI apps the credits map to GPU inference, so the cost is real and immediate. Farmers use scripts, disposable emails, proxies and spoofed browsers, which is why checks that look at one signal at a time miss them.
How do you detect API key reselling?
Look at how keys behave together. Resold keys often come from unrelated accounts yet call the API with the same client, prompts, timing and infrastructure, rotating to stay under per-key limits. Linking the accounts that created the keys and scoring API traffic server side reveals the pool behind them, so you can revoke the set instead of chasing single keys.
Should a SaaS require a credit card for a free trial?
It cuts casual abuse but also cuts honest signups, and determined farmers use stolen or prepaid cards. A middle path works better for most products: let clean signups start without a card, and ask for one only when the signup looks linked to earlier trials or comes from automation. That keeps the funnel open and moves friction onto the abusers.
How can I detect account sharing in a SaaS product?
Compare each session with the account's own history. Sharing shows up as one login used on several devices at once, often in distant locations, or a steady rotation of new devices on a single seat. Flag it for review or ask for a second factor rather than logging users out, since a new laptop or a trip is normal.
Do bot checks slow down developer onboarding?
They do not have to. Checks that run in the background, with no puzzles and an async script that never blocks the page, add nothing to a normal signup. The friction lands only on accounts with risky evidence, such as automation or links to earlier trials. Start in observe-only mode to confirm the effect before enforcing anything.
See who is really on your site.
Run Kavra on your own traffic in observe-only mode. No risk to your customers, and a clear report of the fraud it finds.