| CVE | Vulnerability name | Date | Responsible Security Disclosure by | Vulnerabilities |
|---|---|---|---|---|
|
GHSA-2g94-9x3m-hv37 CVE requested
|
LockoutBleed |
2026-08-04 |
NinjaGPT and xet7![]() ![]() ![]() ![]() ![]() Coordinated disclosure via GitHub Security Advisory GHSA-2g94-9x3m-hv37. |
|
WeKan bundles wekan-accounts-lockout to stop password guessing: by default,
three failed logins lock the account for sixty seconds. It had been doing nothing at all.
Both of its Accounts.validateLoginAttempt hooks decided whether an attempt
was a failure by comparing the error's reason string —
loginInfo.error.reason !== 'Incorrect password' for a known user,
!== 'User not found' for an unknown one — and returned early otherwise.
That reason never arrives. Meteor's accounts-base ships
ambiguousErrorMessages defaulting to true, so its
_handleError rewrites every credential failure — wrong password, no
such user, no password set — to one generic sentence before any
validateLoginAttempt hook runs:
// meteor/accounts-base — accounts_server.js
_handleError = (msg, throwError = true, errorCode = 403) => {
const isErrorAmbiguous = this._options.ambiguousErrorMessages ?? true;
const error = new Meteor.Error(
errorCode,
isErrorAmbiguous
? 'Something went wrong. Please check your credentials.' // <-- the reason the hooks see
: msg, // never 'Incorrect password'
);
...
};
So the literals never matched, both hooks returned early, the failure counter was never incremented, and no account ever locked. The report confirmed it dynamically: the lockout collection stayed empty and unlimited failed logins were accepted — no CAPTCHA, no rate limit, no lockout.
The fix decides from the attempt's structural fields instead of a localized,
Meteor-internal string WeKan does not control (new
packages/wekan-accounts-lockout/src/loginFailureDecision.js): a password
login of a known user that carries an error is a countable failure; a password login with
no matched user that carries an error is a countable unknown-user failure. The one error
deliberately not counted is no-2fa-code — accounts-2fa throws it
after the password already checked out, to ask for the second factor, so it is
the normal first leg of every two-factor sign-in and counting it would lock out
legitimate 2FA users. A wrong second factor (invalid-2fa-code) still
counts, because there the password was already correct.
The accounts-password login path runs a bcrypt comparison (~50 ms) only when the user exists and has a local password. For a missing user — or an LDAP/OIDC-only user with no local password — it throws immediately (~2 ms) with no bcrypt work. The two response-time distributions do not overlap, so an unauthenticated attacker can tell whether any given username or email exists with near-100% reliability, no matter how uniform the error text is. With enumeration free and guessing unthrottled, any account with a weak, guessable or breached password — including an administrator's — could be taken over, a successful login returning a full resume login token.
The standard mitigation is applied: whenever the real path would skip bcrypt, one
dummy bcrypt comparison against a fixed cost-10 hash is performed so the
missing-user path costs about the same as a real check (new
server/lib/loginTimingDefense.js). On the DDP login method a
timing-normalization login handler runs ahead of the built-in password handler, looks the
user up, and — when there is no local password to check — burns the
compensating time before falling through; it never authenticates
(server/loginTimingNormalization.js).
POST /users/login in server/apiAuthRoutes.js checks the password
directly with Accounts._checkPasswordAsync and never runs the DDP lockout
hooks, so on its own it had no brute-force protection at all — and it enumerated by
message, throwing a distinct “User with that username or email address
not found.” for a missing user while a wrong password threw the ambiguous one.
It now fails missing-user and wrong-password with the same uniform error, runs the
same dummy-bcrypt timing equaliser for a missing or password-less account, and throttles
failed attempts per client address (new server/lib/loginAttemptThrottle.js,
default 10 failures / 60 s → 60 s lockout). Only failures count and a success clears
the client's counter, so correct-credential clients are never impeded;
X-Forwarded-For is honoured only when HTTP_FORWARDED_COUNT
declares the proxy depth, so the header cannot be spoofed to dodge the throttle.
Four plain-Node unit suites pin each half: tests/loginFailureDecision.test.cjs
(the ambiguous reason is counted, success still enforces an active lock,
no-2fa-code never locks anyone), tests/loginTimingDefense.test.cjs,
tests/loginAttemptThrottle.test.cjs (the time-injected throttle and the key
resolver), and tests/loginBruteForceEnumerationWiring.test.cjs (the fragile
reason-string guards stay gone and the REST endpoint keeps its uniform error, timing
equaliser and throttle).
Reported by NinjaGPT via GitHub Security
Advisory
GHSA-2g94-9x3m-hv37,
reproduced against v8.35. Fixed at the
upcoming WeKan release.
The DDP half of the timing fix was never loaded. Since WeKan's package.json
sets meteor.mainModule, the server runs only what an import chain from
server/main.js reaches, and server/loginTimingNormalization.js
was not imported, so the web login's response time still told whether a username
exists; the REST login was equalized. It is now loaded, and
tests/serverModulesLoaded.test.cjs fails when any server module that
registers a login handler, method, publication, route, startup hook or job is not
reached. Nothing is refused by this fix, so there is nothing to record in Admin Panel
→ Problems; the lockout itself still records under LockoutBleed.
Fix.