Credential stuffing

Slow down leaked-password attacks on your login.

10 minutes

Leaked username and password lists get tried against your login. Rate limits per IP miss it; counting accounts per visitor does not.

What you will build: a check on the action you protect that allows normal users, asks doubtful ones to verify, and blocks the clear cases. See the credential stuffing overview for the attack in pictures.

Before you start

Finish one Web quickstart (for example JavaScript or React) and one Server quickstart (for example Node.js). The decision code below is TypeScript; every server SDK returns the same event.

How the attack works

  • Combo lists. Millions of leaked pairs, tried in bulk.
  • IP rotation. A new proxy for every attempt.
  • Low and slow. Few tries per account, many accounts.

What to look at

FieldWhy it matters here
velocityDistinct accounts and IPs per visitor.
tlsMismatchAn HTTP client pretending to be a browser.
botAutomated login forms.
  1. 1

    Identify on the action

    Call the browser SDK when the user performs the action, not on page load, and send the eventId with the request. Pass your own user id as linkedId whenever you have one: the account signals count by it.

    TypeScript
    const { eventId } = await trace.get({ linkedId: user?.id, tags: { action: 'credential-stuffing' } });
  2. 2

    Verify on your server

    import { verifyEvent } from '@trace-id/sdk-node';
    
    const event = await verifyEvent(eventId, { secretKey: process.env.TRACE_SECRET_KEY!, endpoint, timeoutMs: 1500 }).catch(() => null);

    Events can be verified for 60 minutes. If verification fails, decide whether to fail open or closed; for most of this flow, fail open.

  3. 3

    Decide

    TypeScript
    const { signals } = await trace.verify(eventId);
    
    if (signals.bot.detected || signals.tlsMismatch.detected) return block();
    if (signals.velocity.detected) return captchaBeforePassword();
    return checkPassword();

    trace.verify() is the embedded call; with a server SDK, use the event from step 2 instead. The thresholds are yours. Trace scores and explains; event.reasons lists which signals added how much, so log it with every decision.

Example decisions

Login attempts from one visitor, illustrative:

AccountIPDecisionWhy
user_118281.4.x.xAllowFirst attempt
user_2093102.33.x.xVerifyNew account, new IP
user_773145.12.x.xBlockvelocity: linkedIdsPerVisitor24h
user_0420188.2.x.xBlockvelocity + tlsMismatch

Roll it out

  1. Ship steps 1 to 3 with logging only: record the decision, block nothing.
  2. After a week, look at who would have been blocked or asked to verify. Shared computers, offices and privacy tools are the usual false positives.
  3. Turn on Verify first, then Block for the clearest cases.
Suggest an edit

Spotted something wrong or unclear? Tell us on GitHub and we will fix it.