Most OTP abuse does not look like an attack. It looks like ordinary traffic: verification requests arriving from many addresses, each one plausible on its own. If your only OTP rate limiting is a per-IP counter, every request gets a fresh budget and your SMS invoice pays for it.
The pattern is usually SMS pumping. Someone triggers codes at volume and the messages are billed to you. The request itself is cheap to send and cheap to disguise, because the IP address is the easiest part of a request to change.
A counter model that holds up needs three axes — phone number, device and network — resolved into one decision you can tune: allow, step up, or block.
Why an IP-only limit lets OTP pumping through
An IP address is not a user. Carrier-grade NAT can put many mobile subscribers behind a single address, an office egresses through one, and a household shares one all day. A tight per-IP cap punishes those real people while an attacker with a proxy pool or a handful of mobile connections walks straight past it.
The economics make it worse. Rotating addresses is cheap; sending a code is not. Any limit that resets when the address changes is a limit the attacker outruns by design, and the damage appears in your provider invoice rather than in your error rate.
The fix is not a higher per-IP threshold. It is counting each request against identities that are harder to rotate.
Count on three axes, not one

Per phone number
The number is what gets monetized, so it deserves its own counter. Track request velocity per number, whether that number ever completed verification, and whether it shows up across many devices. Numbers that request codes repeatedly and never verify are the clearest pattern you can act on. Store the number as a keyed hash and use that hash as the counter key.
Per device
Device identity survives an IP change. Count requests per visitorId instead of per cookie jar, and treat one device requesting codes for several numbers as a stronger signal than a device repeating a single number. Fingerprinting is probabilistic — browsers and networks change — so combine it with network reputation rather than trusting it alone. Our device fingerprinting explainer covers which signals stay stable.
Per network
Add the network the request came from: ASN, subnet, and whether the address belongs to a datacenter, proxy or Tor exit. One subnet spreading requests evenly across many numbers is a different pattern from one number being hammered, so it should hit a different counter. Network signals are also the cheapest way to keep known hosting ranges out of your OTP flow without blocking the carrier traffic your real users share.
Turn the counters into one decision
Three counters produce three opinions. Your policy has to resolve them into one answer, and that answer should not be binary. Prynt Control answers the sms use case with allow, step-up or block, a reason code, and a suggested step such as captcha, verify_email or card_or_passkey. Control never blocks on its own: your code receives the decision and applies it.
const result = await control.decide({
useCase: "sms",
phoneHash,
visitorId,
ip,
});
if (result.decision === "block") {
return res.status(429).json({ reason: result.reason });
}
Step up before you block
Blocking a real user costs a signup; a captcha costs a few seconds. Escalate in order: challenge first, email or passkey verification when the risk repeats, and a block only when the pattern is unambiguous — a number or network already confirmed as abusive. Write down which reason codes map to which step, so support can explain a decision without reading your code.
Run it in mirror mode before you enforce
Any threshold you pick on day one is a guess. Run the counters in observation mode first, let nothing be blocked, and read the distribution of requests per number, device and network. Set the enforcement line above the volume your real users produce, then tighten.

Prynt Control opens with a 14-day mirror period for that reason: you see every decision and reason code while nothing is enforced. Phone numbers and cards are kept only as keyed hashes, so the counters work without retaining raw identifiers.
Log the request id, the reason code and the hashed keys. Three fields are enough to answer “why was this code blocked?” weeks later, and they pair with the abuse patterns in our fraud rings guide.
FAQ

Does rate limiting by IP stop OTP pumping?
Not on its own. It stops naive scripts and it is still worth keeping as a cheap first filter, but shared carrier and office addresses make it inaccurate, and rotating addresses make it easy to evade. Pair it with per-number and per-device counters.
What OTP limit should I set per phone number?
There is no universal number, and any figure published without your traffic behind it would be wrong. Run in mirror mode, measure how often your legitimate users request codes for the same number inside your window, and set the cap above that with headroom. Revisit it whenever a signup flow changes.
Can I rate-limit OTPs without storing phone numbers?
Yes, if the counter key is a hash. Keep a keyed hash of the number, use it as the limit key and as the block-list entry, and rate limiting never needs the raw value. Prynt stores numbers and cards this way, which keeps the retention question small.
Try it free
Prynt Control makes the SMS decision per device, number and network, and returns allow, step-up or block with a reason to your own code. It protects the OTP request; your system still sends the message. Start with Prynt Control and run the first 14 days in mirror mode with nothing blocked.
Try it free
Prynt is device intelligence with a free plan: 5,000 identifications a month, visitor IDs with confidence and bot detection — no credit card. Upgrade to Pro for VPN, proxy, tampering and behavioral signals.