|
|
Message-ID: <CACXwg9Bo11W14JV9+utLgiFxbC1ZqD1HLvBaB1O_RDxGthZHAA@mail.gmail.com> Date: Thu, 19 Nov 2026 08:59:24 -0800 From: Andrey Vasilevsky <anvanster@...il.com> To: oss-security@...ts.openwall.com Subject: CVE-2026-124749: pam-mysql: inverted PAM_DISALLOW_NULL_AUTHTOK gives a pre-auth authentication bypass Hello, pam-MySQL (libpam-mysql, src:pam-mysql) inverts PAM_DISALLOW_NULL_AUTHTOK at both pam_sm_authenticate call sites, so an application that asks PAM to refuse passwordless accounts gets the permissive behaviour instead. On a stock sshd, an unauthenticated network client authenticates as any account whose stored password column is SQL NULL, using any password. The supplied password is never compared on that branch. CVE-2026-124749, assigned by Red Hat off the distros posting. CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H = 9.8. A reading that prices the schema precondition into AC gives 8.1 and I would not argue with it. See the note on the 1.0 line below, where that precondition largely disappears. THE DEFECT, against v0.8.2 - pam_mysql.c:3611, parameter 4, is documented "Whether null authentication tokens should be disallowed". - pam_mysql.c:3874, else branch: vresult = null_inhibited - pam_mysql.c:3877: vresult == 0 -> PAM_MYSQL_ERR_SUCCESS - pam_mysql.c:4708 and :4799 pass !(flags & PAM_DISALLOW_NULL_AUTHTOK), which is 0 exactly when the application set the flag. The password comparison is nested inside "if (row[0] != NULL)" at :3689, so on the NULL branch no comparison happens at all. Any password works, including none. The two pam_sm_chauthtok call sites pass a literal 0 and are correct. That sibling contrast is what makes this a slip rather than an intended convention. WHY A DEFAULT sshd IS THE VECTOR, AND WHY THAT IS THE STING OpenSSH sets PAM_DISALLOW_NULL_AUTHTOK precisely when PermitEmptyPasswords is no, which is the default. So the inversion is reached by the safe configuration and not by the unsafe one: an administrator who has done nothing wrong gets the permissive path, and one who has set PermitEmptyPasswords yes does not. That is also why the obvious workaround is wrong. See MITIGATION below. AFFECTED, AND THE 1.0 LINE IS WORSE Every released version carries it, including the newest upstream commit, which dates from 2024-03-14. 0.8.x: Debian 11/12/13/14/sid, Ubuntu 18.04/20.04/22.04, Devuan, PureOS, Pardus, Deepin, Kali, FreeBSD, Gentoo, Liguros. 1.0.0 betas: Fedora 39-44 and rawhide (beta1), Nix 6 channels (beta2), Mageia 10 and Rosa (beta1), ALT (1.0.0). On 0.8.x the exploitable schema is a nullable password column. On the 1.0 line an empty stored password is exploitable too, because the guard there is "if (row[0] && row0len)" at src/authenticate.c:104, which routes an empty stored password into the same inverted branch at :127. So the NOT NULL DEFAULT '' schema that protects an operator on 0.8.x is the exploitable one on the 1.0 line, and that is the schema a careful author writes. PostfixAdmin declares mailbox.password and admin.password as varchar(255) NOT NULL default ''. Froxlor declares ftp_users.password as varchar(255) NOT NULL with no default. Both shapes are shipped, current schemas. On the distributions carrying the 1.0 betas the schema dependency is therefore largely absent, and AC:L with 9.8 is unambiguous there. MEASURED OVER A REAL sshd Each arm reads 0.8.2-Debian / 1.0.0-beta3 / beta3-plus-polarity-fix. - NULL stored, wrong password: SUCCESS / SUCCESS / DENY - empty stored, wrong password: DENY / SUCCESS / DENY - control, real stored, correct password: SUCCESS / SUCCESS / SUCCESS - control, real stored, wrong password: DENY / DENY / DENY "Wrong password" above means a wrong, non-empty password. THE FIX THAT IS SHIPPING, WHICH IS NOT MINE Ship this one. It is Ferenc Wagner's, and it is better than the patch I originally proposed. old: vresult = null_inhibited; new: vresult = !ctx->nullok || null_inhibited; plus the polarity correction at both pam_sm_authenticate call sites: old: !(flags & PAM_DISALLOW_NULL_AUTHTOK)); new: !!(flags & PAM_DISALLOW_NULL_AUTHTOK)); and a new nullok boolean, default off, registered in both the module-argument and the config-file option tables. The reasoning, which changed my mind about my own patch: before the fix, an application calling with flags=0 does not authenticate a NULL-password account. Correcting the polarity alone would start authenticating one. That is a behaviour change to make in a stable security update, even though it is what the flag is documented to mean. nullok is the same name and the same semantics pam_unix and pam_ldap already use, so an operator who needs the permissive behaviour has an obvious way to ask for it. NULL stored password, any supplied password, reading before / polarity-fix-only / shipping-fix: - called with flags=0: DENY / SUCCESS / DENY, and SUCCESS under the shipping fix only when nullok is enabled - called with PAM_DISALLOW_NULL_AUTHTOK: SUCCESS / DENY / DENY My own patches are in the distros archive and are a fallback, not a recommendation. The 0.8.x one additionally routes an empty stored password into the null branch, so it denies that case by default where the shipping fix leaves 0.8.x behaviour unchanged. On 0.8.x its three hunks have to land together: applying the empty-string hunk without the polarity hunks would make NOT NULL DEFAULT '' columns exploitable in the way the 1.0 line already is. For the 1.0 line the polarity correction is two characters, the same "!!" at src/authenticate.c:287 and :378. No empty-string hunk is wanted there, because the guard already routes both cases to one branch. MITIGATION IF YOU CANNOT PATCH YET Do not set PermitEmptyPasswords yes. It does suppress this defect, because it clears the flag and a cleared flag denies a NULL row, but it also disables OpenSSH's own empty-password guard at auth-passwd.c:93-94, which opens the empty-stored-password case instead. It trades one bypass for another. The supportable mitigation is to stop storing credential-less rows in any table used for authentication: purge them, then add NOT NULL with no empty-string default. One consequence worth stating plainly for operators who are relying, knowingly or not, on the inverted behaviour to refuse NULL rows through an application that permits null tokens. Those deployments become more exposed after a polarity-only fix, not less. The shipping fix avoids that, because nullok defaults off and a NULL stored password is refused regardless of the flag. Either way the real remediation is the rows, not the flag. HOW TO CHECK WHETHER YOU WERE EXPLOITED The caveat first, because it is the important part: OpenSSH's log cannot distinguish a bypass from a legitimate login. Both produce the same line. Accepted password for nullpw from 127.0.0.1 port 39024 ssh2 Accepted password for normalpw from 127.0.0.1 port 39034 ssh2 The first is a bypass of an account whose stored password was NULL. The second is a genuine login. Nothing in either line marks the difference. So the discriminator has to come from the database. Any accepted password login for an account whose stored password was NULL or '' is necessarily a bypass, because no password could legitimately have matched. Start from the credential-less rows: SELECT username FROM users WHERE password IS NULL OR password=''; and correlate those names against "Accepted password for" lines in the auth log. Two things make a naive check miss: 1. Current database state does not prove historical state. A column populated since the login makes the account look safe today. The binary log or a backup from the period is the only reliable source where it matters. 2. pam_mysql's own audit table is off by default, and when enabled it contains a spelling error that defeats the obvious grep. log.enabled defaults to 0 at pam_mysql.c:2897 and is gated at :4331, so a default deployment records nothing there at all. With log.enabled=1 the table does capture the bypass, but the failure string is misspelled AUTHENTICATION FALURE, so a query for FAILURE returns nothing. There is a weak secondary signal in that table: the bypass produced a single SUCCESS (FIRST_PASS) row with no preceding failure, where the genuine login produced a failure then a success, because the first pass tried a token that did not match. I would not build detection on that alone, since it depends on the calling application's pass behaviour, but it is worth knowing when reading an existing table. STATUS Upstream is Nigel Cunningham. He confirmed the defect and approved a fix on 2026-10-26, held the commits for the coordinated date, and pushed them to master at 16:37 UTC today, after the lift. Two commits carry it: 6e2c52eb "CVE Fix", which is the "!!" correction at both pam_sm_authenticate call sites, and 46175520 "Add nullok option.", which brings the null branch to "vresult = !ctx->nullok || null_inhibited", the same shape Debian is shipping. Master is the 1.0 line. The v0.8.x branch has not been updated and its head still dates from 2021, so for the 0.8 line the fix is at present the distribution patch rather than an upstream commit. He also enabled private vulnerability reporting on the repository after my report, so there is now a private channel where there was none. Debian's maintainer is Ferenc Wagner. The trixie security upload carries the fix, the nullok documentation and the CVE reference, and went to security-master on 2026-11-17. Bookworm is Debian LTS rather than stable security, so its update goes through the LTS process and has not been uploaded yet; bookworm users are still exposed as I write this. Bookworm is 0.8.2-2 against trixie's 0.8.2-2.1, the same upstream version with the same patch series, so the fix applies there unchanged. Debian's upload also extends the package's autopkgtest to cover NULL and empty stored passwords under both flag values, which is the regression test this defect should have had from the start. I have not checked the other affected distributions. The 1.0-line ones have no fixed release: today's master commits carry no tag, and the newest tag is still 1.0.0-beta3. Fedora, Nix, Mageia, Rosa and ALT ship those betas, where the exploitable schema is the common one, and the two-character polarity correction above applies to them directly. TIMELINE - 2026-10-23 found, reproduced end to end on Debian trixie - 2026-10-24 reported privately to upstream and to Debian - 2026-10-26 upstream confirmed and approved a fix - 2026-11-08 posted to distros, embargo opened - 2026-11-09 CVE-2026-124749 assigned by Red Hat - 2026-11-17 Debian trixie upload to security-master - 2026-11-19 14:00 UTC embargo lift; upstream pushed to master at 16:37 UTC; this posting One correction to that timeline rather than leave it to be noticed. I told upstream and Debian on 10-24 that I proposed 2026-11-07 as the disclosure date, then failed to open the distros thread in time for that date to mean anything, so it passed unused and the embargo clock started on 11-08 instead. Nothing was published and nothing leaked; the coordination sat in a private thread two weeks longer than it should have, and that was my fault rather than anyone else's. CREDIT Nigel Cunningham, upstream, for confirming quickly and then sitting on finished patches for eleven days while I sorted out a gap that was mine. Ferenc Wagner, Debian, for the fix that is actually shipping and for the regression test. Salvatore Bonaccorso and the Debian Security Team for the coordination. Rafael Pandini of Red Hat Product Security for the CVE assignment. A full write-up, both of my patches, and a container-based reproducer that installs the distribution package and compiles nothing of the module are available on request. I used AI assistance to find and write up this issue. Every code citation and every measured arm was verified by hand. Thank you, Andrey Vasilevsky
Powered by blists - more mailing lists
Please check out the Open Source Software Security Wiki, which is counterpart to this mailing list.
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.