Users should manually update to the latest version now
UPDATED A series of flaws in Tailscale, an open source mesh virtual private network (VPN) software, could allow attackers to stage remote code execution (RCE) attacks against VPN nodes.
Tailscale depends on multiple services. The main process, called tailscaled, does the work of connecting nodes and sending/receiving packets.
There is a separate process that provides a user interface and a tray icon to configure and monitor the services. This front-end interface communicates with the tailscaled service through an HTTP API called LocalAPI.
From DNS rebinding to control plane takeover
If a malicious website tries to send a JavaScript command to the Tailscale LocalAPI, the browser’s Same-Origin policy will prevent it.
However, according to the findings of security researcher Emily Trau and Jamie McClymont, if the attacker manages to perform a DNS rebinding attack on the Tailscale node, they will be able to map their malicious domain to the local IP and send arbitrary commands to the LocalAPI.
“Rebinding is a bug with very niche applicability (HTTP services listening on private networks with no explicit authentication), usually discussed in the context of IoT devices,” McClymont told The Daily Swig. “It’s the type of thing there are talks about it at hacker conferences and such, and yet I’ve never encountered a situation where it’s exploitable during a pentest job.”
The LocalAPI does not authenticate client requests aside from verifying that they’re coming from the same user that is running the Tailscale GUI.
The malicious website can exploit this feature to change the Tailscale “control plane” to an arbitrary server. The “control plane” is the server that stores the public keys of the VPN nodes (also called the tailnet).
In a tailspin
As the tailnet administrator, the attacker can now enable Taildrop, a feature that allows users to send files between their devices on a Tailscale network.
Using Taildrop, the attacker can then send an arbitrary executable to the victim’s desktop without marking it as originating from the web, which means Tailscale will be able to launch it without requiring user interaction.
To execute the payload, the attacker can use another feature of the control plane, which demands the Tailscale node to reauthenticate itself when trying to perform a privileged action. The re-authentication prompt includes an address that is forwarded to the GUI and runs it in the browser.
To run the file, the attacker will need to have its full path, which requires knowledge of the victim’s username. To obtain the victim’s username, the attacker can prompt for an SMB path through the Tailscale network. This will send the Windows username to the attacker-controlled tailnet server.
“For the Windows RCE chain, you just need to click a link to an attacker-controlled webpage, or for the malicious Javascript to be served up to you some other way (e.g. malicious JS embedded in an ad on a legit site, or an XSS bug in legit site being exploited to add the malicious code),” McClymont said.
If you were running a stable Tailscale version, the exploit will lie dormant until you next restart Tailscale or reboot your machine, at which point the RCE happens.
“You could argue this is still zero interaction since Windows Update will automatically reboot without interaction eventually,” McClymont said. “If you had the unstable pre-release Tailscale version from right before we found the bugs, we could trigger the exploit immediately without waiting for a reboot.”
A catch on DNS rebinding
A recent modification to the Same-Origin policy forbids rebinding a site that was hosted on a public IP to a private IP space. This prevents the attacker from rebinding an internet-hosted malicious website to a local IP address.
But it is still applicable if the attacker is on the same network as the victim. Also, the Firefox browser does not apply the private network address restriction, which makes it vulnerable to internet-hosted attacks.
Moreover, Trau found that PeerAPI, another Tailscale component, runs on the 100.100.100.100 IP and was vulnerable to rebinding, which would give the attacker another pathway to LocalAPI.
Also, if the attacker sends multiple files to the victim’s device through Taildrop, some of them will fail to reach their destination and will remain in a temporary location that is reachable through web calls without private network access restrictions.
Trau published a proof-of-concept video of the attack. Windows machines are especially vulnerable to different variations of the attack. Other operating systems can also be exploited under special circumstances.
The issues have been solved in the latest version of Tailscale. Since Tailscale does not automatically update itself, users should make sure they are running v1.32.3 or later.
“If you’re running HTTP services over Tailscale which rely on Tailscale for authentication (they don’t have their own login page etc.), you need to harden them against rebinding attacks, either by allowlisting Host headers or by running those services only on HTTPS,” McClymont said.
This article has been updated to include comment from researchers.
RECOMMENDED Intel disputes seriousness of Data Centre Manager authentication flaw
