Where Cynopta Protects Your App
The same human-likelihood signal can support controls across these automated-abuse scenarios. Your app or backend decides whether to allow, step up, throttle, delay, or reject.
Account Takeover (ATO) Mitigation
Flags automated login attempts so your app can ask for a second factor or refuse them.
Read more
Passkeys protect the moment of login, so attackers now aim at what happens around it: stolen session tokens, remote access, coached victims and account recovery. UK Finance reported a 21 percent rise in mobile banking fraud cases in the UK in 2025. UK Finance
A high score at login and a low score at payment can point to a session that started with a person and continued without one. Your policy decides what follows, such as a passkey prompt or a manual review. We have not measured how often the score shows that change in one session.
- A coached victim is a real person. When a criminal on a call talks a customer through a transfer, the score correctly shows a person at the phone.
- The score is not an authentication factor and not identity. It does not replace passkeys or one-time passcodes, and it does not tell you who holds the phone.
- Cynopta does not provide a score that your servers can check for themselves today. For a payee change or a withdrawal, pair the score with a Play Integrity result that your server verifies. Play Integrity confirms a genuine app on a genuine phone; the score asks whether a person is using it.
List the actions in your app that move money or control after login, such as a new payee, a changed phone number or email address, account recovery and payments. For each one, check whether your app asks for anything more than the login it already accepted. Those actions are the first places to read the score.
Stolen Credential Validation
Gives your login a human signal, so scripted runs against leaked password lists are refused before they reach your backend.
Read more
Credential stuffing reaches a mobile app in two ways. Some attackers run your app itself, on emulators, virtual phones in the cloud or racks of real phones, so their requests look like your customers' requests. The SDK scores those sessions. Other tools call your login servers directly and never open the app, so rate limits and bot management on your servers still carry that traffic.
If your login shows a challenge, keep it, and use the score to decide which sessions see it. A challenge alone separates people from software less well than it did: researchers at ETH Zurich reported that their software solved 100 percent of Google's "reCAPTCHA v2" image challenges. ETH Zurich
- Logins on your website never reach the SDK. Keep your challenge on web login.
- A person paid to tap a phone, and software that imitates human touch well, are harder cases. The score may treat them as a person.
Pick the flow in your app that shows the most challenges, often login or sign-up. Count how many challenges people see there in one week, and how many they pass. That count is your baseline for a watch-only trial, in which your app records the score next to each challenge result and takes no action on it.
Fake Registration Defense
Flags automated sign-ups so mass-created accounts never join your user base or distort your metrics.
Read more
In October 2025, police in Latvia seized 1,200 SIM boxes and 40,000 active SIM cards from one service. Europol The Hacker News reported that the service "enabled the creation of more than 49 million online accounts." The Hacker News A SIM box holds many SIM cards, so one computer can receive verification codes for many numbers.
Farms also rent cloud phones, which run in a data center and are controlled over the internet. Geekflare says that one cloud phone service charges from USD 0.007 a minute for each phone. Geekflare
On Android, your app reads the score at sign-up without a network call. The score checks for emulators, virtual phones and cloud phones, for signs that software is driving the screen, and for physical signs typical of phones in a rack, such as a phone that never moves and is always on power.
- People who create extra accounts by hand score as people, because a person is using the phone.
- The SDK scores each phone. It does not link phones or accounts today, so it does not show that many accounts belong to one farm.
- The part of the score that comes from the phone's daily routine takes from 3 days to 3 months to become meaningful, so at sign-up on a new install it has had little time.
If your Android app uses Play Integrity, check whether it reads the "recent device activity" level. Google says that it shows "whether a device has made an anomalously high volume of requests recently, which could indicate automated traffic". Google If your app does not read it yet, record it at sign-up.
SMS Toll Fraud Protection
Lets your app skip the verification text when no human is present, so SMS pumping costs you nothing.
Read more
The UK National Cyber Security Centre describes the fraud: a bot creates many fake accounts, each one makes your service send a one-time passcode, and a rogue party in the mobile network takes the messages, never delivers them and shares the profit. UK NCSC Nobody needs to read the code. The money is made when the message is sent, and you pay for it on your next messaging invoice.
Juniper Research estimated that this artificially inflated traffic cost businesses USD 2.1 billion in 2023. Juniper Research
- Scripts that call the endpoint that sends your codes, without opening your app, never reach the SDK. Server controls still matter: accept only valid number formats, limit codes for each user, and do not send codes to countries or number types that you do not need. UK NCSC
- A real person who requests codes by hand scores as a person.
Compare the codes that your app sends with the codes that people enter, by country and by hour. The NCSC calls this the conversion rate. Read the score in watch-only mode for a few weeks, act on it, and measure again. If the codes sent fall and the conversion rate rises, some of the fake requests came through your app. If they do not fall, the requests probably reach your server without your app, and your server controls need the work.
Carding Attack Prevention
Gives your checkout a human signal, so card testing is refused before the payment gateway is called and the bot learns nothing from declines.
Read more
Visa said in May 2024 that enumeration attacks cause USD 1.1 billion in fraud losses a year. Visa Its acquirer monitoring program, VAMP, sets two enumeration thresholds: enumerated attempts reaching 20 percent of all authorization attempts, or a count of 300,000. Visa
In a native app, card details meet your gateway mainly at checkouts for physical goods and services, at wallet top-ups and on screens where a person saves a card. Your policy then decides: call the gateway, ask for a stronger check such as 3-D Secure, or not call the gateway at all.
- Card tests sent to your website or straight to your payment servers do not pass through the SDK.
- The score says nothing about whether a card is stolen. A real person who tests stolen cards by hand scores as a person.
- The score does not change how Visa counts enumeration. At most, it keeps some of the card tests that come through your app away from the gateway.
Ask your acquirer or payment provider for your enumeration figures under VAMP. Then split your authorization attempts by channel: your app, your website and direct calls to your payment server. If many card tests come through the app, read the score in watch-only mode at checkout for a few weeks, and compare declined attempts by score.
Discount and Inventory Abuse Prevention
Adds a human check to carts, queues and raffle entries, so limited stock and flash sales go to real customers.
Read more
The FTC says that one broker used SIM boxes to receive Ticketmaster's verification codes and, for one Eras Tour show, used 49 accounts to buy 273 tickets where the limit was six. FTC A purchase limit counts items for each account, so it says little when one buyer controls many accounts.
On Android, your app can read the score when someone creates an account, joins a queue, checks out or transfers a ticket. Your policy can then give the full purchase limit to accounts on phones that look like a real person is using them.
- People paid to buy by hand score as people, even when they use SIM boxes and virtual payment cards.
- The SDK does not link accounts or phones to each other today, so it cannot tell you that many accounts belong to one buyer.
- Sales through your website never reach the SDK.
List each point in your app where a text message code is the main proof that a buyer is a separate person. For each one, write down what else your app checks, and what it would see if one buyer ran thousands of accounts through it.
Fake Review Prevention
Attaches a human signal to review submission, so coordinated rating campaigns are filtered before they move your scores.
Read more
The FTC rule on consumer reviews, in effect since 21 October 2024, bans fake reviews and fake signs of social media influence, "such as followers or views generated by a bot or hijacked account". FTC In the UK, fake reviews have been banned since 6 April 2025, and platforms must "take steps to prevent and remove the publication of fake reviews." GOV.UK
Your app can read the score at account creation, to slow down or refuse accounts from phones that look automated, and at review submission, to store it with each review as one more input to your moderation.
- People paid to write reviews, and sellers who swap reviews, write by hand on their own phones. The score correctly shows a person.
- Do not label reviews as "verified human" because of the score. The score is a likelihood, and a label that presents it as proof would mislead the people who read your reviews.
Take the reviews that your moderators removed in the last three months. Count how many came from accounts created shortly before the review. If many did, account creation is a good place to add a check. If most came from older accounts, paid people or review swaps are the more likely cause, and your moderation team is the better tool.
Spam Mitigation
Scores the sending device at send time, so your app can throttle or block automated floods in comments, forums and chat.
Read more
Match Group told a US congressional committee that one threat actor can create "hundreds of thousands of accounts using automation" in a coordinated spam attack. It banned nearly 36 million spam and inauthentic accounts worldwide in 2024. US Congress JEC
Your app can read the score when a sender account is created, set a stricter rate limit for phones that look automated, and read it again when a quiet account starts sending at volume. The score comes from signals such as the phone's motion and touches on the screen, not from message text, so it works the same way in chats that are end-to-end encrypted.
- Workers in scam centers who type by hand on real phones score as people, even when they work for a scam operation.
- Once a conversation moves to another app, your app cannot score it.
Take the accounts that your team banned for scams in the last three months. Check how old each account was and how many messages it sent before the ban. If many were new accounts that sent many messages quickly, account creation and the first messages are where a score on the phone can help.
Data Scraping Protection
Flags automated harvesting through your app, so your APIs can require an acceptable score before serving prices, content or user data.
Read more
OWASP lists scraping among its automated threats to web applications, and its other names for it include "Comparative shopping" and "LLM training bot". OWASP The Wikimedia Foundation found that "at least 65% of this resource-consuming traffic" to its website came from bots, while bots made about 35% of all pageviews. Wikimedia
A scraper reaches the data in your app in two ways: it calls your servers directly, or it runs your app on many phones and taps through product pages like a customer. For the second, your app can read the score before it asks your servers for prices or content.
- Scripts that call your API without your app never reach the SDK.
- The score does not tell a search engine from a scraper. Some automation helps you, such as a partner that reads your catalog under contract, and which automation you allow is your decision.
- A person who copies your prices by hand on a real phone can score as a person.
API Abuse and Infrastructure Cost Control
Lets your app refuse automated, resource-heavy API traffic inside the app, before it reaches your backend and your cloud bill.
Read more
OWASP's API Security Top 10 lists "Unrestricted Resource Consumption" as the fourth risk. It says that the risk "can also lead to operational costs increase", for example with services "paid for per request, such as sending emails/SMS/phone calls". OWASP
Crawling operators now use residential proxies and spoof real web browsers, the Wikimedia Foundation wrote in March 2026, so a limit for each IP address helps less than it did. Wikimedia A bot farm on real phones goes further: its requests come from mobile networks and from your own app. Your app can read the score and turn low-scoring requests away before your backend spends a cycle.
- Cynopta does not protect your service from a distributed denial-of-service (DDoS) attack at the network layer. Keep your CDN and DDoS providers for that job.
- Scripts that call your API directly never reach the SDK. Rate limits and spending limits on your servers still carry that traffic.
List the API endpoints that return your prices, stock or profiles, and the paid services that your app calls for each request. Check each one against the limits that OWASP lists for the resource consumption risk, such as the number of records in one response and a spending limit with each provider.
Bonus and Referral Abuse
One person or farm opening many accounts to claim sign-up bonuses, referral rewards or free bets. A human signal keeps rewards for people.
Read more
The pattern repeats across industries. Singapore police said that four men had allegedly colluded "to create fake Grab accounts and manipulate the platform to submit fraudulent claims". Singapore Police Force Linea's airdrop analysis initially flagged 516,960 of 1,297,203 eligible addresses as Sybil addresses, which one person or group controls while posing as separate users. The Block In one week in 2025, PUBG Mobile banned 315,940 accounts, 60 percent of them at the game's entry tier. India Today Gaming
On Android, read the score at sign-up and again when your app pays a reward, redeems points or allows a first withdrawal. The part of the score that comes from the phone's daily routine takes from 3 days to 3 months to become meaningful. So it is not ready for a coupon used on the day of sign-up, but it can be ready by the first withdrawal.
- A real person who opens extra accounts by hand, or rents out a verified account, scores as a person.
- The score says whether a person or automation is at the phone, not whether that person is honest. Collusion and false refund claims need other checks.
- Farming that happens only on a blockchain, with wallets that never use your app, does not reach the SDK.
Take the accounts that received your welcome coupon or referral reward last month. Count how many placed one order and did not return, and how many of those shared a delivery address or a payment card with other accounts. If that group is large, record the score in watch-only mode at sign-up and when you pay the reward, before your app acts on it.
Mule and Farmed Accounts
Accounts opened and run from phone farms to move money through exchanges and banks. A human signal shows which accounts no person operates.
Read more
A US indictment describes "two particular facilities staffed with 1,250 mobile phones that controlled 76,000 accounts on a popular social media platform." US DOJ In India, the government had blocked about 303,000 IMEI numbers by January 2026, so the authorities now act against the phone itself, not only the SIM card and the bank account. PIB
A real phone does not prove that a person is there. Google says that abusers "generate valid attestation results from real devices and provide them to bots". Google In the Philippines, Circular 1213 requires measures that prohibit "unauthorized scripts or automation tools", with behavior analysis and bot detection. BSP
On Android, your app can read the score at sign-up, login on a new phone, adding a payee, payment and withdrawal. Store it next to your own record of each event, with the time when your app read it.
- A real person who rents out their own account scores as a person. The score does not show that an account is rented.
- In a coerced payment, the customer is a real person using their own phone, so the score correctly shows a person. Payee checks, warnings and delays fit that case.
- A score does not make your firm compliant with Circular 1213 or any other rule. Your regulator decides that.
Choose one event that your regulator cares about, such as a withdrawal or a login on a new phone. Take ten recent cases and check whether your records show how each one was driven, by a person or by automation, and what your app knew at the time.
Agents Acting for Customers
A human signal, so your policy can admit the agents your customers send and stop the rest.
Read more
When ByteDance's Doubao assistant began tapping through other apps on a phone in December 2025, large Chinese apps reacted with rules for the whole app or account: WeChat logged users out, and two bank apps told users to turn off the AI tool. Alipay's reaction was narrower: it stopped the task at the password step. Yicai Global
Agent signatures and payment mandates tell you which agent is at your door. They do not tell you whether anyone sent it, or whether a person was at the phone. On Android, the score checks whether software, not a finger, is driving the screen, and whether someone unlocked the screen shortly before the session began. That answers whether someone was at the phone, not who it was or whether they agreed to what happens next. Your app can set a low bar to track a parcel and a high bar to change a payee.
Agents also learn to look human. Researchers found that an agent with system-level access to the phone can forge touch and motion together, and six behavioral detectors accepted 77.5 percent of its events as human. arXiv A score that combines many physical signals is harder to fool, but not impossible to fool, and we have not measured how the Cynopta score holds up against these attacks.
- Cynopta does not tell you which agent acted, for whom, or within what limits. Signatures and mandates are designed for those questions.
- A hijacked assistant and an honest one both look automated. Whether an automated action is wanted is your decision.
- Agent payments that run through card networks and payment APIs do not open your app, so there is no session to score.
List each flow in your app. Mark whether an assistant may complete it, prepare it for a person to confirm, or must leave it to a person. Then compare the list with what your app does today when it sees automation.
Each scenario runs on the same mechanism: one SDK, one human-likelihood score - used alone, or feeding the fraud engine, bot manager, and authentication stack you already run.
Find Out Who Is Behind Your Traffic
Free up to 10,000 assessments a month. Business from $8 a month.