Every free credit on an AI platform is compute someone pays for. Score each sign-up before it gets credits, stage the credits and keep the card for accounts that raise a signal.
AI free-trial abuse prevention starts before credits are issued. Score each signup with device intelligence, email, phone, network and account-link signals, stage the credits and ask for a card or identity only when the risk score requires it. A card at the door turns away real developers and still lets operations with stolen cards through.
Free trial abuse is opening accounts, usually many of them, to claim a platform's free trial or free credits again and again. This guide is part of our guide to token fraud.
At a glance
Stripe said on 10 March 2026 that 7.4% of customer sign-ups at AI companies on its network were "implicated in suspected multi-account abuse". It is a vendor's own figure, with no method published.
Stripe's later "one in six" figures use four different wordings. None of them matches the 7.4%, so this guide does not use them as a level.
Stripe wrote on 15 September 2026 that "AI companies are increasingly exposed to fraud before a transaction takes place."
One AI video company with a free trial logged more than 105,000 automated sign-up attempts from 251 IP addresses in a month, Okta Threat Intelligence reported in August 2026.
Stytch, which supplied Kilo Code's device fingerprinting, says Kilo Code asked for a card only from sign-ups the check flagged. Everyone else claimed free credits without one.
What is free trial abuse on an AI platform?
Free trial abuse on an AI platform is one person or operation opening many accounts to claim free credits that were meant for one customer each. The credits buy inference, so each extra account spends GPU time the platform pays for.
It usually takes one of three shapes:
Accounts opened to claim free credits. An operation signs up again and again, then spends the credits, sells the accounts or routes paying traffic through them.
Accounts paid for with stolen cards. Where credits need a card, the operation uses stolen numbers, often after testing them on your form. Our guide to card testing and chargebacks covers that side.
Banned users on new accounts. Someone you removed comes back under a new name. Our ban evasion guide covers how to recognise them.
Why does free credit abuse cost AI platforms more than other trials?
A free month of ordinary software costs its maker almost nothing, but a free credit on an AI platform is paid GPU time. Credits claimed in bulk also rarely stay with the account that claimed them.
"When a bad actor spins up five accounts instead of one, they consume five times the compute resources," Stripe wrote in its March 2026 analysis.
The credits then become supply for resale. Stytch's case study of Kilo Code describes attackers who claimed free credits on thousands of accounts, "then reselling those accounts or routing LLM traffic through them for profit." Okta Threat Intelligence found resale services crediting cloud sign-up bonuses for their low prices and judged that "fraudulent registration is one of the reasons" the market thrives. Our token resale guide follows those accounts into the API relays that sell them on.
What does a multi-account operation look like?
An operation runs scripts that open accounts in bursts, each with a throwaway email address, a fresh IP address and often a rented phone number. The details change on every account, while the person and setup behind them stay the same.
Okta described an AI video company with a free trial. "Between June 1, 2026, and July 1, 2026, logs show over 105,000 brute force attempts to sign up from 251 distinct IPs likely connected to bots." The sign-up list contained disposable email domains, and Okta found a seller offering ready-made accounts on the same service for US$40 each.
Kilo Code's team saw "huge bursts of new signups from unusual IP ranges, all redeeming free credits at once", according to Stytch. Requiring Google sign-in did not stop them, and its CAPTCHA worked only briefly.
Does blocking disposable email addresses stop it?
Blocking disposable email domains stops the cheapest sign-ups at almost no cost to real customers. It does not stop an operation that owns its own domain or uses ordinary mailboxes.
Block lists are never complete. The open-source disposable-email-domains list held 9,913 domains on 19 October 2026, and its maintainers had last updated it the day before.
A domain you own defeats a list completely. In 2023, the security firm Checkmarx showed that OpenAI's sign-up flow could be beaten with a catch-all email account on a private domain, which accepts mail at any address, SecurityWeek reported. Each new address looked unique to OpenAI's check.
Castle, which sells sign-up protection, advises "Filtering disposable emails selectively, rather than blocking free email providers entirely" in its 2025 post on AI free tiers. Treat a disposable domain as one reason in a score.
Do VoIP and burner phone numbers get past phone verification?
Phone verification raises the cost of each account, but operations rent numbers that receive the text for them. A text code proves someone can read a message at that number, and says nothing about whether the number has opened accounts before.
Castle found verification texts from Mistral AI, Poe, Claude and OpenAI arriving at free temporary phone numbers, and noted that operations can also rent numbers "by the day, just for the account creation process". In May 2026, Castle published an open list of the 1,000 most active disposable phone numbers it sees in abuse, updated daily.
Stripe's own advice on multi-account sign-up abuse says phone verification "can be bypassed with Voice over Internet Protocol (VoIP) numbers". It recommends "phone confirmation with a real carrier look-up" for sign-ups in the middle band. A carrier look-up such as Twilio's line type check separates a "nonFixedVoip" number, "obtained online without requiring a physical device", from a mobile line.
Checkmarx also showed that OpenAI's phone check counted variations of one number, such as added leading zeros, as different numbers, so one phone could claim credits many times. Normalise every number before you compare it, and score a VoIP number as one reason instead of rejecting it, because some real customers use office VoIP lines.
Should a free trial require a credit card?
A card at sign-up cuts casual repeat sign-ups and costs you real customers. It does little against an operation that already holds stolen, prepaid or virtual cards, and it gives card testers a free way to check numbers.
Stripe's own guide to preventing free trial abuse says a required card "ties that trial to a real financial identity". It also lists cards among the last controls to add, because "Each step reduces abuse but slows conversion." Stripe's free trial abuse documentation names "prepaid or virtual cards" and "the same card across multiple accounts" as common abuse patterns, so the card is a signal to score, not proof of a new customer.
Kilo Code weighed the same choice. "Most developers weren't willing to enter payment info just to test a new coding tool," Stytch wrote. Stytch says Kilo Code then asked for a card only when a sign-up looked risky. Kilo Code's own sign-up copy in June 2025 asked every new user for a card hold, as our case on free tiers cut by abuse records, and Stytch's story is undated.
A card requirement also invites stolen cards. In August 2026, dax, who builds OpenCode, said on X that people kept opening accounts for a launch offer and filing chargebacks, often on stolen cards.
What does checking identity history add?
Identity history asks whether the email address, phone number or username has a past outside your platform. A real person's details usually turn up on other sites over years, while details created for one sign-up usually have no history at all.
Email and phone checks prove only that someone controls an address or number. In the words of Stripe's guide, "Neither email nor phone confirmation on their own establishes that an incoming account is distinct from accounts a business has already seen."
History answers a different question. A new address with no presence anywhere, or details that match an account you banned, are reasons to hold back credits. A long-standing address is a reason to let the account through, even on a shared network.
What can Stripe Radar see, and what does it miss?
Stripe Radar learns from payments across Stripe's network, and Stripe extended it to score sign-ups and account activity. Its strongest view is still of cards, while much signup abuse happens before any card appears.
Stripe announced on 29 April 2026 that "Radar now evaluates sign-ups and usage in real time, drawing on signals from across the Stripe network." Its customer abuse page says Radar "works across your business, whether or not you use Stripe for payments." Its free trial control is narrower. The documentation says Radar must know "which payment method collections represent a free trial" and asks you to turn on "Radar on payment methods saved for future use".
Stripe's 15 September 2026 post explains why that matters. Stripe says AI companies now need to assess the customer as well as the payment because fraud starts before a transaction. The same post says attempted multi-account abuse at AI subscription companies rose 40% from January to June 2026.
If your free credits need no card, a check that starts from the payment method has little to judge at the moment the credits go out. That is the honest reason to run a sign-up check before Radar, not instead of it.
How common is free trial abuse at AI companies?
Stripe has published the only network-wide figures, and they measure different things. The 7.4% figure has a stated population, and the "one in six" figures come in four wordings that each count something else.
Figure, in the source's words
Date and source
What it counts
7.4% of customer sign-ups "implicated in suspected multi-account abuse"
10 March 2026, Stripe blog
Sign-ups at AI companies on Stripe. No window or method given, and the abuse is "suspected".
"one in six attempted sign-ups is made by a bad actor"
29 April 2026, Stripe newsroom
Attempts, blocked ones included, with no definition of a bad actor.
"token thieves now account for one in every six new customer signups"
The April figure with "attempted" dropped and multi-account abuse named as the cause.
"1 in 6 registrations at AI companies are from bad actors already known to Stripe"
Undated, Stripe Radar product page
Registrations matched to Stripe's own history of abuse.
40% rise in attempted multi-account abuse
15 September 2026, Stripe blog
Change at AI subscription companies, January to June 2026.
The one-in-six figures run more than twice as high as the 7.4%, because they count blocked attempts, people Stripe already distrusts or theft in general. We cite the 7.4% for the level and the 40% for the trend. Both come from one vendor's network with no published method.
How do you score a sign-up before giving credits?
Score each sign-up on its email, phone, network and history together, and act on bands of the score. A combined score catches an operation that passes each single check, because the operation's details rarely all look clean at once.
A workable set of bands:
Low risk. Give the full credits.
Middle band. Give fewer credits, or ask for a carrier-checked phone number before the rest.
High risk. Ask for a card or identity before any credits, as Kilo Code did.
Matches an account you removed. Block, or send to your review queue.
Real customers trip single checks all the time. Fingerprint, which sells device checks, warns in its post on ban evasion that "legitimate users often share IPs". Students on a university network and people behind privacy relays need the other signals to carry them, so one reason should never block on its own.
Put the card last, after the checks that cost real customers nothing
When sign-up abuse spikes, the first instinct is to put a card in front of the free credit. We think that is the wrong first move, because it loses the developer who will not give a card and barely slows the operation that has a supply of them.
The order we recommend:
Score the sign-up before the account exists, on its email, phone, network and identity history.
Stage the credits, so an account that gets through cannot spend a month's allowance on day one.
Watch the first day of usage. US agencies told AI companies in advisory AA26-251A to watch for "new subscriptions immediately at maximum usage as opposed to gradual AI adoption".
Ask for a card or identity only from accounts that raised a signal. Kilo Code put its card check there, and Anthropic does the same with identity. Anthropic's September 2026 report says that when it sees signs of abuse such as resale, "our systems can require users to verify their identity to retain access."
What to check on your platform
You can run each of these against your own data this week:
Disposable domains. Join last month's sign-ups to an open disposable-domain list and count how many credits those accounts used.
Phone line types. Run a carrier look-up on a sample of verified numbers and count the numbers a carrier marks as virtual.
Shared networks. Group sign-ups by network range and day, and look at any range with dozens of new accounts.
Day-one usage. List accounts that used most of their credits within 24 hours of signing up.
Returning details. Compare new sign-ups with the card, phone and email patterns of accounts you banned in the last six months.
Credit cost. Multiply credits per account by the accounts you later linked to abuse and by your compute cost per credit.
Appeals. Count how many blocked users appealed and won, to see what your checks cost real customers.
Where NoResponse fits
Our sign-up check is one API call you make when someone creates an account. It looks up the email address, username or phone number across about 4,000 sites and returns a risk score, the reasons and a decision in about three seconds, before the account's first request.
Usage fingerprints run in your gateway and learn how each account uses your platform. They recognise banned users on new accounts and accounts that resell your tokens. Relay intelligence tracks 70+ relay and resale services and shows which of your accounts supply them. During a wave of sign-ups, incident support gives you hands-on help.
Book a demo and we'll show you how many of last month's sign-ups we would have flagged.
Stripe, which sells a fraud product, says 7.4% of sign-ups at AI companies on its network were implicated in suspected multi-account abuse in March 2026. It also says attempted multi-account abuse rose 40% from January to June 2026. No lab or agency has published an independent figure.
Should a free trial require a credit card?
Not as the first check. A card turns away real developers who won't share one, and stolen, prepaid and virtual cards still get through. Score the sign-up first and ask for a card only when the account raises a signal.
How do I prevent users from creating multiple accounts?
Score every sign-up on its email, phone, network and identity history before it gets credits, and compare it with the accounts you have already removed. Then watch the first day of usage, because accounts run by one operation behave alike from the start.
