The [Qubes U2F Proxy] is a secure proxy intended to make use of U2F two-factor authentication devices with web browsers without exposing the browser to the full USB stack, not unlike the [USB keyboard and mouse proxies][USB] implemented in Qubes.
[U2F], which stands for "Universal 2nd Factor", is a framework for authentication using hardware devices (U2F tokens) as "second factors", i.e. *what you have* as opposed to *what you know*, like a passphrase.
This additional control provides [good protection][krebs] in cases in which the passphrase is stolen (e.g. by phishing or keylogging).
While passphrase compromise may not be obvious to the user, a physical device that cannot be duplicated must be stolen to be used outside of the owner's control.
Nonetheless, it is important to note at the outset that U2F cannot guarantee security when the host system is compromised (e.g. a malware-infected operating system under an adversary's control).
The U2F specification defines protocols for multiple layers from USB to the browser API, and the whole stack is intended to be used with web applications (most commonly websites) in browsers.
In most cases, tokens are USB dongles.
The protocol is very simple, allowing the devices to store very little state inside (so the tokens may be reasonably cheap) while simultaneously authenticating a virtually unlimited number of services (so each person needs only one token, not one token per application).
The user interface is usually limited to a single LED and a button that is pressed to confirm each transaction, so the devices themselves are also easy to use.
Currently, the most common form of two-step authentication consists of a numeric code that the user manually types into a web application.
These codes are typically generated by an app on the user's smartphone or sent via SMS.
By now, it is well-known that this form of two-step authentication is vulnerable to phishing and man-in-the-middle attacks due to the fact that the application requesting the two-step authentication code is typically not itself authenticated by the user.
(In other words, users can accidentally give their codes to attackers because they do not always know who is really requesting the code.) In the U2F model, by contrast, the browser ensures that the token receives valid information about the web application requesting authentication, so the token knows which application it is authenticating (for details, see [here][u2f-details]).
Nonetheless, [some attacks are still possible][wired] even with U2F (more on this below).
In a conventional setup, web browsers and the USB stack (to which the U2F token is connected) are all running in the same monolithic OS.
Since the U2F model assumes that the browser is trustworthy, any browser in the OS is able to access any key stored on the U2F token.
The user has no way to know which keys have been accessed by which browsers for which services.
If any of the browsers are compromised, it should be assumed that all of the token's keys have been compromised.
(This problem can be mitigated, however, if the U2F device has a special display to show the user what's being authenticated.) Moreover, since the USB stack is in the same monolithic OS, the system is vulnerable to attacks like [BadUSB].
In Qubes OS, by contrast, it is possible to securely compartmentalise the browser in one qube and the USB stack in another so that they are always kept separate from each other.
The Qubes U2F Proxy then allows the token connected to the USB stack in one qube to communicate with the browser in a separate qube.
We operate under the assumption that the USB stack is untrusted from the point of view of the browser and also that the browser is not to be trusted blindly by the token.
Therefore, the token is never in the same qube as the browser.
Our proxy forwards only the data necessary to actually perform the authentication, leaving all unnecessary data out, so it won't become a vector of attack.
This is depicted in the diagram below (click for full size).
The `vault` qube with a dashed line in the bottom portion of the diagram depicts future work in which we plan to implement the Qubes U2F Proxy with a software token in an isolated qube rather than a physical hardware token.
This is similar to the manner in which [Split GPG] allows us to emulate the smart card model without physical smart cards.
One very important assumption of U2F is that the browser verifies every request sent to the U2F token --- in particular, that the web application sending an authentication request matches the application that would be authenticated by answering that request (in order to prevent, e.g., a phishing site from sending an authentication request for your bank's site).
With the WebUSB feature in Chrome, however, a malicious website can [bypass][wired] this safeguard by connecting directly to the token instead of using the browser's U2F API.
The Qubes U2F Proxy also prevents this class of attacks by implementing an additional verification layer.
This verification layer allows you to enforce, for example, that the web browser in your `twitter` qube can only access the U2F key associated with `https://twitter.com`.
This means that if anything in your `twitter` qube were compromised --- the browser or even the OS itself --- it would still not be able to access the U2F keys on your token for any other websites or services, like your email and bank accounts.
This is another significant security advantage over monolithic systems.
(For details and instructions, see the [Advanced usage] section below.)
For even more protection, you can combine this with the [Qubes firewall] to ensure, for example, that the browser in your `banking` qube accesses only one website (your bank's website).
By configuring the Qubes firewall to prevent your `banking` qube from accessing any other websites, you reduce the risk of another website compromising the browser in an attempt to bypass U2F authentication.
The above assumes a `work` qube in which you would like to enable u2f. Repeat the `qvm-service` command for all qubes that should have the proxy enabled. Alternatively, you can add `qubes-u2f-proxy` in VM settings -> Services in the Qube Manager of each qube you would like to enable the service.
If you are using Qubes 4.0, you can further compartmentalise your U2F keys by restricting each qube's access to specific keys.
For example, you could make it so that your `twitter` qube (and, therefore, all web browsers in your `twitter` qube) can access only the key on your U2F token for `https://twitter.com`, regardless of whether any of the web browsers in your `twitter` qube or the `twitter` qube itself are compromised.
If your `twitter` qube makes an authentication request for your bank website, it will be denied at the Qubes policy level.
Next, empty the contents of `/etc/qubes-rpc/policy/u2f.Authenticate` so that it is a blank file.
Do not delete the file itself.
(If you do, the default file will be recreated the next time you update, so it will no longer be empty.) Finally, follow your web application's instructions to enroll your token and use it as usual.
(This enrollment process depends on the web application and is in no way specific to Qubes U2F.)
The default model is to allow a qube to access all and only the keys that were enrolled by that qube.
For example, if your `banking` qube enrolls your banking key, and your `twitter` qube enrolls your Twitter key, then your `banking` qube will have access to your banking key but not your Twitter key, and your `twitter` qube will have access to your Twitter key but not your banking key.
The large number of possible combinations of TemplateVM (Fedora 27, 28; Debian 8, 9) and browser (multiple Google Chrome versions, multiple Chromium versions, multiple Firefox versions) made it impractical for us to test every combination that users are likely to attempt with the Qubes U2F Proxy.
