A security researcher has published a report claiming that flaws in Tailscale SSH could let an attacker who has already compromised one permitted device climb to full root access on other servers in the same private network, without stealing a password or a key. The report, titled Tailworm and published on October 25 by Erica Windisch of Cyberdione Labs, says the chain moved from device to device in under a second per hop in laboratory tests.

The report lists 11 findings across Tailscale's SSH server, its serve and app-connector features, the DERP relay client, and headscale, the self-hosted Tailscale server. Windisch tested Tailscale 1.104.2 and headscale 0.29.5 between October 22 and 24. The caveat matters: these are lab results, Tailscale has not confirmed the specific claims, and no bulletin, CVE, or fix had been published at report time. Tailscale says it is working with security specialists on the issues.

How the SSH flaw is said to work

Tailscale SSH runs inside tailscaled, a background program with full root power. When a user may forward Unix sockets, the local channels used by services such as Docker, the program opens or replaces those channels with root rights. A local forwarding rule could thus reach root-only sockets; in the lab an ordinary session gained root and read the /etc/shadow password file almost instantly. A remote forwarding rule could replace a root-owned socket for code execution, and a timing flaw in a path check could let a rogue node join as a trusted server.

Tailscale does not supply the first foothold: that could be phishing, a vulnerable web app, a poisoned package, or stolen logins. But every device receives a full list of its peers, so an attacker can see every target, and broad access rules let the chain run.

Connectors, serve and headscale findings

The experimental conn25 app connector allegedly checks only a destination's address family, not the configured app destinations, so an entitled device could reach any destination, including cloud metadata holding server credentials and internal Kubernetes services. On the serve side, a TCP forward to the local API socket could let a remote peer write to the local API and take over a node.

Headscale warnings: unauthenticated registrations could push out pending device approvals, a hostile relay could force a client to read up to 1.34 gigabytes, and a memory-exhaustion loop could block new devices from joining. On headscale 0.29.5, SSH forwarding cannot be switched off, so every SSH grant should be treated as equivalent to root access until a fix lands.

What administrators should do now

With no patch available yet, the report's interim advice starts with access rules: replace broad SSH grants with per-pair rules, set forwarding to denied wherever possible, and remove unneeded connector permissions and serve forwards to Unix sockets. Teams can switch Tailscale SSH off on devices that do not need it, but should confirm another way to reach the device first. Turning on node-key expiry and enabling SSH check mode, which adds a login step but does not fix the escalation, are also recommended.

If a device is compromised, the report says to delete it at the control plane rather than simply reimaging the host, because keys already copied by an attacker survive a rebuild. In July, Tailscale fixed TS-2026-009, a dash-username lookup flaw, in version 1.98.9. Some commentators dispute how severe the new findings are, and the results remain unconfirmed lab work rather than observed breaches, but teams running Tailscale SSH should review their rules now and watch Tailscale's bulletins.