TailWorm Tailscale SSH flaws: patch to 1.104.4
Tailscale bulletins TS-2026-014 to 017 cover the TailWorm SSH and Serve flaws. Find out which nodes were exposed and how to move each one to 1.104.4 or newer.
What TailWorm is, and the short answer
TailWorm is the name of a report on flaws in Tailscale SSH and Tailscale Serve. The fix is Tailscale client 1.104.4 or newer on every node. Tailscale published four security bulletins for it, TS-2026-014 to TS-2026-017. Update the nodes that run Tailscale SSH first. If you cannot update a node today, set TS_SSH_DISABLE_FORWARDING=1 in the environment of its tailscaled service and restart the service.
Tailscale's own pages spell the report's name "Tailworm". The bulletin IDs and version numbers in this guide come from Tailscale's own bulletins and changelog. If you are new to the product, read what Tailscale is and how a tailnet works first. The rest of this page assumes you know what a node and a tailnet policy are.
The four TailWorm bulletins, and which nodes each one hits
The bulletins page dates all four bulletins Oct 26, 2026. Two are about Tailscale SSH. Two are about Tailscale Serve.
- TS-2026-014, Tailscale SSH Unix socket forwarding. Affected: users who have Tailscale SSH enabled and whose access policy also allows local or remote forwarding. A peer could use
ssh -L,ssh -Ror a race in the path allowlist to reach Unix sockets that only root may open. The bulletin names the Docker socket and thetailscaledlocal API as examples. - TS-2026-015, Tailscale SSH session memory exhaustion. Affected: nodes running the Tailscale SSH server. The attacker needs an authenticated SSH session that the policy permits. The result is an out-of-memory kill of
tailscaled, which disrupts all Tailscale connections on that node. The bulletin lists no workaround. - TS-2026-016, Serve Unix backends. A Serve backend on a Unix socket could reach privileged sockets.
- TS-2026-017, Serve HTTP/2 shared context. A shared-context flaw in how Serve handled HTTP/2.
All four are fixed in Tailscale 1.104.4 or newer. The changelog entry for 1.104.4 states the main SSH fix directly: "Tailscale SSH performs Unix socket forwarding on behalf of the authenticated user rather than as root." That sentence also explains TS-2026-014. Forwarding ran as root, so a peer that logged in as an ordinary user still had its forwarded connections opened with root's permissions. The Docker socket matters here because a process that can talk to it can start a container with the host's filesystem mounted. That is close to root on the box.
The same changelog entry says SSH session memory is now bounded and released on disconnect. For Serve, it restricts TCP forwards and fixes a capability leak.
Were only Tailscale SSH nodes exposed?
For TS-2026-014 and TS-2026-015, yes. Both bulletins name Tailscale SSH as the condition. For TS-2026-016 and TS-2026-017, no, because those are Serve bugs. The bulletins have no "not affected" section. So do not treat a node as safe because it lacks Tailscale SSH. Update every node.
Affected versions, platforms and CVE numbers
The bulletins give no first affected version. They say only that the fix is in 1.104.4 or newer, so treat any client older than 1.104.4 as unpatched. Version 1.104.3 was internal only, and 1.104.2 (Oct 21) carries no SSH fixes.
The bulletins name no operating system, and the changelog lists 1.104.4 for all platforms. The Tailscale SSH server itself only runs on Linux and on the open-source tailscale and tailscaled command-line build for macOS, according to the Tailscale SSH docs. Look at those nodes first for the two SSH bulletins.
The bulletins and the blog post list no CVE (Common Vulnerabilities and Exposures) identifiers. Cite the TS-2026-0xx IDs when you file a ticket or answer an auditor. A scanner that matches on CVE numbers will not know about these four bulletins. That is one reason to check your server for known CVEs and also track vendor bulletins by hand.
The threat model in plain words
Tailscale's blog post of Oct 28, 2026 sets out the threat model. The flaws do not let an unauthenticated attacker connect to arbitrary Tailscale SSH servers. An attacker first needs a foothold on a peer that is already authorized in your tailnet, with SSH access explicitly permitted in your policy file. They also need the conditions that each bug requires, such as forwarding being allowed. The post also says: "They don't describe self-replicating malware spreading automatically between machines." So the name is not a description. Nothing spreads on its own.
In practice, the risk is a peer that you do not fully trust, or one that was compromised some other way, and that your policy lets in over Tailscale SSH. A laptop with malware or a contractor's device are typical examples. The attacker uses that peer to reach a node it can already log in to. Then it moves from an ordinary SSH user to sockets that only root may open. For the wider picture of what the Tailscale design protects, see whether Tailscale is secure and what its control plane can see.
The timeline, as Tailscale tells it
These dates come from Tailscale's blog post and its other pages. News coverage links the researcher's own disclosure at tailworm.com, but that site would not load for this guide. Check it yourself for the researcher's full list of findings and the versions they tested.
- Oct 23: concerns start to circulate, per Tailscale.
- Oct 24: Tailscale receives the report in the morning. The researcher publishes it two hours later.
- Oct 26: the bulletins page dates TS-2026-014 to 017, and Tailscale posts an investigation update with the forwarding workaround.
- Oct 28: Tailscale 1.104.4 ships, and the blog post goes up.
- Oct 30: headscale 0.29.7 ships as a security release.
How to check your Tailscale version
Run these on each node:
tailscale version
tailscale version --daemon
tailscale version --upstreamtailscale version prints the version of the command-line client. --daemon also prints the version of the running tailscaled daemon. That is the one that matters, because the SSH server runs inside the daemon. --upstream prints the latest release published on pkgs.tailscale.com, so you can compare.
The healthy result is a daemon version of 1.104.4 or higher. A client that says 1.104.4 next to a daemon that still says 1.104.2 means the package updated but the old daemon is still running. Restart it with sudo systemctl restart tailscaled and run tailscale version --daemon again.
For a whole tailnet, open the Machines page in the admin console. It has a Version column, so you can find every node below 1.104.4 in one view.
How to update Tailscale on Linux through the package repo
If you installed Tailscale from its package repository, update it like any other package. On Ubuntu and Debian:
sudo apt-get update
sudo apt-get install tailscale
tailscale version --daemonOn RHEL 8, CentOS Stream 9 and Fedora, Tailscale's update docs use sudo dnf update tailscale. CentOS 7 and Amazon Linux 2 use yum, and openSUSE uses zypper.
If apt-get install tailscale does not offer 1.104.4 or newer, check that the Tailscale repository is configured. On Ubuntu 24.04 (codename noble) the repo setup is two lines:
curl -fsSL https://pkgs.tailscale.com/stable/ubuntu/noble.noarmor.gpg | sudo tee /usr/share/keyrings/tailscale-archive-keyring.gpg >/dev/null
curl -fsSL https://pkgs.tailscale.com/stable/ubuntu/noble.tailscale-keyring.list | sudo tee /etc/apt/sources.list.d/tailscale.listThen run the apt-get update and apt-get install lines again. If apt reports a key or repository error at this step, fixing Tailscale install errors on Ubuntu covers the common messages.
Write "1.104.4 or newer" in your notes and tickets, not "latest". The stable channel moves on. "Latest" in a ticket written today means a different version next month.
How to update with the built-in tailscale update command
Tailscale has its own updater, tailscale update. It exists in version 1.36 and later. It works on Windows and on some Linux distributions, including Ubuntu and Debian, but not on every distro.
sudo tailscale update --dry-run
sudo tailscale update --yes
tailscale version --daemon--dry-run shows what the updater would do and changes nothing. --yes skips the interactive prompt, which you want in a script. --track picks the stable, release-candidate or unstable channel. --version asks for one exact version, and you cannot combine it with --track.
To keep nodes current from now on, turn on automatic updates per node:
sudo tailscale set --auto-updateNew tailnets have auto-updates on by default. Tailnets created before February 2024 must turn them on. Tailscale's docs say auto-updates are not supported on Synology or on Arch-based distributions, so update those nodes by hand.
How to tell whether Tailscale SSH is on
Tailscale SSH needs two things. The node must run the SSH server, and the tailnet policy must have an ssh rule that lets someone in. Check both.
On the node. Tailscale SSH is switched on per node with tailscale set --ssh (or tailscale up --ssh). The quickest documented check is the admin console: you can filter the machine list by whether a device has Tailscale SSH enabled. That gives you the exact list of nodes the two SSH bulletins are about. The server needs Tailscale 1.24 or later, and Tailscale's docs say Synology and QNAP do not support it. Any platform can still connect as an SSH client.
If a node does not need the SSH server, turn it off. A node that does not run the server does not meet the affected conditions of TS-2026-014 or TS-2026-015.
sudo tailscale set --ssh=falseTailscale's docs warn you to make sure you have another way to SSH to the device first. If Tailscale SSH was your only way in, this command locks you out.
In the policy. Open the tailnet policy file in the admin console and find the ssh section. Each rule has an action (accept or check), a src (who connects), a dst (which nodes) and users (which local accounts they may log in as). Read each rule as one sentence: "these sources may SSH into these nodes as these users." Every src in that list is a peer that, if compromised, meets the precondition in Tailscale's threat model.
Stop-gap: disable forwarding with TS_SSH_DISABLE_FORWARDING
For a node you cannot update yet, the TS-2026-014 bulletin gives a workaround. Set TS_SSH_DISABLE_FORWARDING=1 in the tailscaled service environment and restart the service. This turns off SSH forwarding and leaves the rest of Tailscale SSH working. Tailscale's investigation update said it "mitigates several of the more serious findings". The bulletins list it for TS-2026-014 only, and TS-2026-015 has no workaround. Updating is still the real fix.
The bulletin does not say which file to put the variable in. On a systemd host, a drop-in override is the standard way to add an environment variable to a service:
sudo systemctl edit tailscaledIn the editor, add these two lines in the space it offers, then save:
[Service]
Environment=TS_SSH_DISABLE_FORWARDING=1Then restart the service and check it:
sudo systemctl restart tailscaled
systemctl show tailscaled --property=EnvironmentThe second command should print a line containing TS_SSH_DISABLE_FORWARDING=1. If it prints only Environment=, the override was not saved. Run systemctl cat tailscaled to see which files systemd actually loaded.
Do the restart from a console or from a normal, non-Tailscale SSH session. Tailscale SSH sessions run inside tailscaled (which is why TS-2026-015 can take the daemon down), so restarting the daemon ends them.
Harden the ssh rules the way Tailscale recommends
Beyond the update, Tailscale's blog post asks you to limit Tailscale SSH access to the users and devices that need it, and to consider check mode for sensitive systems. The forwarding advice from the bulletin still applies: do not allow forwarding you do not need, or set the variable above. The blog post gives no rule syntax. The example below is one way to apply that advice. It is not text from the bulletin.
"ssh": [
{
"action": "check",
"src": ["group:admins"],
"dst": ["tag:prod"],
"users": ["autogroup:nonroot"]
}
]What each choice does:
"src": ["group:admins"]names a group instead of every member of the tailnet. A rule'ssrccannot be*, and neither can itsdst."dst": ["tag:prod"]limits the rule to tagged servers, so a new laptop does not become an SSH target by accident.autogroup:nonrootallows any local user that is notroot. Root logins then need their own, separate rule."action": "check"turns on check mode. The user must sign in again with their identity provider before the connection opens, on the device that starts the connection. The default check period is 12 hours, andcheckPeriodchanges it.
Check mode is optional and off by default. A checkPeriod of always asks every time, and Tailscale's docs warn that this can cause unexpected behavior with automation tools such as Ansible. Keep a separate accept rule, scoped to the automation host's tag, for those jobs. Tailscale's SSH docs also say Tailscale SSH may not suit multi-user machines. Weigh that on a box shared by several people.
The ssh section only governs Tailscale SSH on port 22. It does not replace your network rules, so also review the ports and firewall rules Tailscale needs on the same nodes.
If you run headscale
Headscale is a self-hosted control server. It does not run the SSH server or Serve, because those live in tailscaled on each node. So upgrading headscale alone patches none of your nodes. Update every client to 1.104.4 or newer.
Headscale did ship its own security release. Version 0.29.7, dated 2026-10-30, fixes findings from Anthropic's oss-scanner project and the Tailworm report. Its changelog sets the minimum supported Tailscale client to v1.80.0. It also urges users, especially Tailscale SSH users, to move clients to 1.104.4 or newer. For the server upgrade itself, follow running headscale as a self-hosted Tailscale control server.
FAQ
Is my node affected by TailWorm if I never turned on Tailscale SSH?
Not by the two SSH bulletins, TS-2026-014 and TS-2026-015. Both require the Tailscale SSH server. The other two, TS-2026-016 and TS-2026-017, are flaws in Tailscale Serve and do not depend on SSH. The bulletins publish no "not affected" list, so update every node to Tailscale 1.104.4 or newer.
What CVE numbers do the TailWorm flaws have?
Tailscale's bulletins and its blog post list no CVE identifiers. The IDs to cite are TS-2026-014, TS-2026-015, TS-2026-016 and TS-2026-017 from Tailscale's security bulletins page. A vulnerability scanner that only matches CVE numbers will not flag these. Check the version on each node with tailscale version --daemon instead.
Is TailWorm a worm that spreads between machines?
No. Tailscale's blog post says the findings do not describe self-replicating malware. Exploiting them needs an existing foothold on a peer that is already authorized in your tailnet, and SSH access to the target permitted in your policy file. The fix is to update to 1.104.4 or newer and to limit which peers your ssh rules let in.
Does upgrading headscale fix the TailWorm flaws?
No. The Tailscale SSH and Serve fixes are in the Tailscale client, version 1.104.4, which runs on each node. Headscale 0.29.7 is a separate security release for the control server, and its changelog urges updating clients to 1.104.4 or newer. Upgrade both.