Alpine Linux and the OpenSSH team shipped emergency fixes on November 5 for a password check that could be beaten by typing too much. On some musl-based Linux systems, accounts that are supposed to never accept a password, including key-only users and the default alpine user on official cloud images, would accept almost any password longer than 256 bytes. A threat-intelligence firm says the first exploitation attempt hit its sensors less than nine hours after the bug went public.
Key points
- CVE-2026-123158, nicknamed “sshole,” lets an attacker log in to password-less accounts on musl-based Linux systems running OpenSSH without PAM, such as a default Alpine server, by sending an overlong password.
- OpenSSH 10.6p2 and patched Alpine packages for every branch from 3.16 to edge shipped on November 5, but the running sshd stays vulnerable until it is restarted.
- Previdian says its sensors logged 19 attempts from 12 IP addresses through November 7 and that one attacker got root on its Alpine honeypot.
- Debian, Ubuntu, Fedora and other glibc systems, PAM-based setups and servers with password logins turned off are not affected.
What broke, in plain English
When you log in to a Linux server with a password, sshd (the OpenSSH server) takes what you typed, hashes it with the C library’s crypt() function, and compares the result with the hash stored in /etc/shadow. If they match, you’re in.
Accounts that should never take a password store a single asterisk, *, instead of a hash. An asterisk isn’t a valid hash, so in theory nothing can ever match it. That convention goes back decades; the bug’s finder, Özgün Kültekin of Wakefield, traces it on his write-up at sshhole.com to a 1980 BSD password file.
The trouble is what musl, the small C library that Alpine Linux is built on, does when hashing fails. Instead of returning nothing, its crypt() returns a single *. And musl’s SHA-256 and SHA-512 password hashing refuses inputs longer than 256 bytes, a limit added in 2012 to stop attackers from tying up a server with giant passwords. Put those together and a password that’s too long produces a “hash” of *, which matches the * stored for a password-less account. The NVD record describes it as a “misinterpretation of whether a string is intended to be a password hash” and gives it a CVSS score of 8.1 out of 10 (High).
There’s one more condition. According to Alpine’s advisory, the machine needs at least one other local account with a real $5$ or $6$ password hash. That’s because of an older OpenSSH fix: to stop attackers from timing logins to learn which usernames exist (CVE-2016-6210), sshd borrows another account’s salt and hashes the attempt anyway. Kültekin’s write-up argues that this is what sends the doomed * account down the hashing path in the first place. Alpine’s version: “One user with a password is enough: it makes every account without one reachable.”
Alpine didn’t soften the impact. “An attacker who can reach your SSH port can log in without knowing any password. It takes one connection, works every time, and needs no retries,” its security team wrote.
Who is actually exposed
This is not a “patch every Linux box on Earth” bug, and it’s worth being precise. Per Alpine’s advisory and the OpenSSH 10.6p2 release notes, a system has to meet all of these conditions:
- It runs OpenSSH’s sshd on Linux with musl libc and without PAM (Pluggable Authentication Modules, the usual Linux login framework). Alpine’s default
openssh-serverpackage is built that way, andopenssh-server-pamis affected too ifUsePAM nois set. - Password logins are enabled, which is OpenSSH’s default.
- Some account has
*as its password field: a key-only user created withadduserand no password, users pulled in from a directory service through musl-nscd, or root, which has shipped asroot:*since about Alpine 3.16. (OpenSSH’s defaultPermitRootLogin prohibit-passwordstill blocks direct root logins.) - At least one other local account has a real password hash.
The nastiest case is the cloud. Alpine says its official tiny-cloud images from 3.18.11 through 3.24.3 ship an alpine user with a * password and passwordless doas (Alpine’s lightweight sudo). Those machines are safe at first boot. The moment an admin sets any real password, say a console password for emergencies, that alpine account becomes a password-free door, and doas turns it into root. Images configured through cloud-init are spared because cloud-init writes PasswordAuthentication no.
Kültekin also names OpenWrt, Buildroot and Yocto, the toolkits behind a lot of routers and embedded gear, plus “a lot of the docker images you’ve pulled.” Two caveats keep that from being as scary as it sounds. OpenWrt uses Dropbear for SSH by default, and Dropbear isn’t affected. And Alpine says containers that don’t run sshd inside them aren’t exposed. The SecurityWeek and BleepingComputer write-ups list the same preconditions.
Not affected: Debian, Ubuntu, Fedora, RHEL and other glibc-based distributions, anything that checks passwords through PAM, OpenBSD itself, and any server with PasswordAuthentication no.
From disclosure to first attack, in under a day
The timeline below comes from OpenSSH, Alpine and the event log on Previdian’s tracking page, converted to Pacific time.
- Nov 5, 7:46 a.m. PT: the CVE is published, the same day OpenSSH releases 10.6p2 and Alpine pushes fixed packages for branches 3.16 through edge.
- Nov 5, 9:30 a.m. PT: Previdian logs the first public proof-of-concept code. Its page lists seven proof-of-concept or scanner listings created within about a day, which it warns are unverified.
- Nov 5, 4:05 p.m. PT: Previdian’s sensors record the first exploitation attempt, a little over eight hours after publication.
- Nov 6: Previdian founder Ryan Dewhurst tells Help Net Security and The Hacker News that the company “confirmed successful root access on a Previdian honeypot running real Alpine Linux and OpenSSH,” and that the attacker ran reconnaissance commands like
id,unameandhostname. At that point, he counted eight attempts from one Japan-based address. - Through Nov 7: Previdian’s page, updated November 8, now shows 19 attempts from 12 unique IP addresses in 8 countries, caught by 4 of its sensors.
The catch: those are one vendor’s numbers from its own sensor network, and Previdian says plainly that its sensor counts “do not show that the target was vulnerable or that the attempt succeeded.” The honeypot root login is a company-reported finding we can’t independently check. Nineteen attempts is also tiny next to the background noise any SSH port sees in a day. What matters isn’t the volume; it’s that someone turned the advisory into working logins on the same day it came out.
One other detail matters for anyone keeping a patch tracker. As of CISA’s most recent Known Exploited Vulnerabilities catalog release on November 6, which we checked against the agency’s data feed on November 8, the bug isn’t listed. Some aggregator sites already show it as a CISA KEV entry; the one actually saying “exploited” is Previdian’s own feed. If your team only moves on CISA’s list, this one hasn’t reached it yet.
Updating isn’t the same as fixing
Alpine’s first instruction is two steps, and the second is easy to skip: update openssh, then restart sshd. “Upgrading the package alone does not replace the running daemon,” the advisory says. A server that ran apk upgrade on Thursday but never restarted sshd is still running the old binary.
Here’s another trap: the fixed version numbers don’t all say 10.6p2. Alpine backported the fix to each branch, so 3.24 is fixed in openssh 10.3_p1-r2, 3.23 in 10.2_p1-r1, 3.22 in 10.0_p1-r11, 3.21 in 9.9_p2-r1, and so on down to 9.0_p1-r6 on 3.16. Alpine point releases 3.21.10, 3.22.8, 3.23.8 and 3.24.4 include the fixed package. A scanner that only checks for “OpenSSH older than 10.6p2” will flag patched Alpine boxes and may confuse the people reading its report.
If you can’t update, both Alpine and OpenSSH say the stopgap is to turn off password logins with PasswordAuthentication no and restart sshd. OpenSSH developer Damien Miller put it more bluntly on Mastodon, telling anyone running musl-based Linux without PAM to “stop reading and set PasswordAuthentication=no or disable SSH for these systems IMMEDIATELY,” as quoted by Help Net Security. For anyone wondering whether they were already hit, Alpine and Orca Security say to search auth logs for “Accepted password for” on accounts that should never accept a password. Orca’s advice is to treat any such host as compromised.
Whose bug is it, anyway?
Nobody in this story wrote obviously wrong code, which is what makes it interesting. OpenSSH’s release notes say the team considers “the root cause of this bug to be outside sshd, but it is exploitable via sshd and can be avoided there.” Musl’s choice to return * on failure was meant as a safety feature, so programs that never check for errors wouldn’t crash. The 256-byte cap was a defense against denial-of-service attacks. Alpine’s switch of root from ! to * in 3.16 fixed a different problem: Kültekin’s history says the earlier lock broke SSH key logins on systems without PAM after Alpine’s 2019 fix for a passwordless-root bug in its Docker images (CVE-2019-5021).
Each layer made a reasonable call. The bug lives in the gap between them, and Kültekin’s account explains why it lasted: OpenSSH is mainly tested on OpenBSD and glibc, while most distributions route logins through PAM, which never calls crypt() for a * account. Musl-without-PAM is a small corner of the matrix, and that corner happens to include a lot of cloud instances. By his count the condition has existed since 2016.
The AI angle, without the hype
Kültekin says he found the flaw by building sshd in every combination he could think of (glibc and musl, with and without PAM) and throwing about 17,600 odd logins at the builds. Every build agreed on every login except musl without PAM. Then, on the same page, comes the aside: “oh did i say ‘i’? sorry. i had agent swarms.”
That fits a warning OpenSSH has put at the top of every release note since August. The team says it receives “a large number of security bug reports, many of which are findings from AI models or made with AI assistance,” and that it has seen AI-found bugs “subsequently independently discovered by a different researcher.” Its conclusion: “adversaries who do not report bugs to OSS projects are likely to be able to discover these bugs too,” so it now ships fixes more often instead of batching them.
Our read: sshole is a good example of what automated bug-hunting is actually good at. This wasn’t a clever new exploit technique. It was a boring differential test (same input, many builds, find the one that disagrees) run at a scale a single person wouldn’t bother with. The same approach works for attackers, and the same-day exploitation shows how short the window between “patch available” and “patch needed” has become. We’ll be back on this beat; we covered OpenAI’s Critical cyber rating for GPT-6 two days ago.
The takeaway
If you run Alpine or any musl-based Linux with OpenSSH, the fix is cheap: update, restart sshd, and turn off password logins if you don’t use them. Everyone else can relax, but should look closely at their container images and embedded gear before deciding they’re in the clear.
Questions still open on this one:
- Does CISA add it to the KEV catalog? That would set federal patch deadlines and push it up a lot of corporate to-do lists that currently skip it.
- How many exposed hosts actually meet all the conditions? Nobody has published a credible internet-wide count of Alpine sshd servers with password logins enabled and a hashed account on board, and that number is what decides whether this stays a niche incident.
- Does musl change
crypt()‘s failure value? OpenSSH worked around the problem on its side. Any other program on musl that comparescrypt()output the same way could have the same blind spot. - Do downstream images get rebuilt? Cloud marketplace images and third-party container bases built on vulnerable Alpine releases need fresh builds, not just a patched package upstream.
Hype in. Signal out.
Image: A screen showing the OpenBSD Secure Shell server (OpenSSH) starting during a Linux boot, 2017. Photo by Hpott via Wikimedia Commons, CC BY-SA 4.0. Cropped and resized for this post.








