Tang is a server for binding data to network presence.
This sounds fancy, but the concept is simple. You have some data, but you only want it to be available when the system containing the data is on a certain, usually secure, network. This is where Tang comes in.
First, the client gets a list of the Tang server's advertised asymmetric keys. This can happen online by a simple HTTP GET. Alternatively, since the keys are asymmetric, the public key list can be distributed out of band.
Second, the client uses one of these public keys to generate a unique, cryptographically strong encryption key. The data is then encrypted using this key. Once the data is encrypted, the key is discarded. Some small metadata is produced as part of this operation which the client should store in a convenient location. This process of encrypting data is the provisioning step.
Third, when the client is ready to access its data, it simply loads the metadata produced in the provisioning step and performs an HTTP POST in order to recover the encryption key. This process is the recovery step.
Tang provides an easy and secure alternative to key escrows.
Before Tang, automated decryption usually took the form of generating a key, encrypting data with it and then storing the key in a remote server. This remote server is called a key escrow.
The concept of key escrow is simple, but managing it can be complex.
Key escrows are stateful by nature. And since they store live data (the encryption keys), they must be surrounded by a sophisticated backup policy. This backup policy also needs to be carefully secured, otherwise improper access to the keys could be obtained. Further, since keys are transferred over the wire, typically SSL/TLS is used. SSL/TLS is a large protocol, with a corresponding large attack surface; resulting in attacks like Heartbleed. Even further, escrows require a comprehensive authentication policy. Without this any user on the network can fetch any key. Often this is deployed using X.509 certificates, which bring their own complexity.
In contrast, Tang is stateless and doesn't require TLS or authentication. Tang also has limited knowledge. Unlike escrows, where the server has knowledge of every key ever used, Tang never sees a single client key. Tang never gains any identifying information from the client.
| Escrow | Tang | |
|---|---|---|
| Stateless | No | Yes |
| X.509 | Required | Optional |
| SSL/TLS | Required | Optional |
| Authentication | Required | Optional |
| Anonymous | No | Yes |
Tang requires a few other software libraries:
- http-parser >= 2.8.0 - https://github.com/nodejs/http-parser
- systemd - https://github.com/systemd/systemd
- jose >= 8 - https://github.com/latchset/jose
Tang is packaged for Fedora. This package should be used as it contains additional settings (such as SETGID directories) out of the box. To install it:
$ sudo dnf install tang
If you really want to build from source on Fedora, you will need the following packages:
- http-parser -
http-parser-devel - systemd -
systemd - jose -
jose,libjose-devel - curl - curl (only needed for running tests)
Tang is also capable of running on devices without systemd even for example OpenWrt (see: this PR). Instead of using systemd for socket activation you can use another daemon for spawning services like xinetd. Without systemd watching Tang's database directory you will have to make sure that key cache is generated properly.
An example of configuration file for Tang using xinetd can be found in the
units/ directory.
Tang is also available as a Docker Container.
Care should be taken to ensure that, when deploying in a container cluster, that the Tang keys are not stored on the same physical medium that you wish to protect.
Building Tang is fairly straightforward:
$ autoreconf -if
$ ./configure --prefix=/usr --libdir=/usr/lib64
$ make
$ sudo make install
You can even run the tests if you'd like:
$ make check
Once installed, starting a Tang server is simple:
$ sudo systemctl enable tangd.socket --now
This command will enable Tang for startup at boot and will additionally start it immediately. During the first startup, your initial signing and exchange keys will be generated automatically.
That's it! You're up and running!
It is important to periodically rotate your keys. This is a simple three step process. In this example, we will rotate only a signing key; but all key types should be rotated.
First, generate the new keys (see jose documentation for more options):
$ sudo jose jwk gen -i '{"alg":"ES512"}' -o /var/db/tang/newsig.jwk
$ sudo jose jwk gen -i '{"alg":"ECMR"}' -o /var/db/tang/newexc.jwk
Second, disable advertisement of the previous key:
$ sudo mv /var/db/tang/oldsig.jwk /var/db/tang/.oldsig.jwk
Third, after some reasonable period of time you may delete the old keys. You should only delete the old keys when you are sure that no client require them anymore. You have been warned.
Tang relies on the JSON Object Signing and Encryption (JOSE) standards. All messages in the Tang protocol are valid JOSE objects. Because of this, you can easily write your own trivial Tang clients using off-the-shelf JOSE libraries and/or command-line utilities. However, this also implies that comprehending the Tang protocol will require a basic understanding of JOSE objects.
All Tang messages are transported using a simple HTTP REST API.
| Method | Path | Operation |
|---|---|---|
GET |
/adv |
Fetch public keys |
GET |
/adv/{kid} |
Fetch public keys using specified signing key |
POST |
/rec/{kid} |
Perform recovery using specified exchange key |
The advertisement reply message contains a JWS-signed JWKSet.
The (outer) JWS contains signatures using all of the advertised signing JWKs.
The (inner) JWKSet contains all of the advertised public JWKs. This includes all advertised signing, encryption and exchange JWKs.
Typically, a client will perform "Trust On First Use" in order to trust the server's advertisement. However, once the client trusts at least one signing JWK, further advertisements can be requested using that signing JWK. This allows clients to upgrade their chain of trust.
Tang implements the McCallum-Relyea exchange as described below.
The basic idea of a McCallum-Relyea exchange is that the client performs an ECDH key exchange in order to produce the binding key, but then discards its own private key so that the Tang server is the only party that can reconstitute the binding key. Additionally, a third, ephemeral key is used to blind the client's public key and the binding key so that only the client can unblind them. In short, blinding makes the recovery request and response indistinguishable from random to both eavesdroppers and the Tang server itself.
The POST request and reply bodies are JWK objects.
The client selects one of the Tang server's exchange keys (sJWK; identified
by the use of deriveKey in the sJWK's key_ops attribute). The client
generates a new (random) JWK (cJWK). The client performs its half of a
standard ECDH exchange producing dJWK which it uses to encrypt the data.
Afterwards, it discards dJWK and the private key from cJWK.
The client then stores cJWK for later use in the recovery step. Generally
speaking, the client may also store other data, such as the URL of the Tang
server or the trusted advertisement signing keys.
Expressed mathematically (capital = private key):
s = g * S # sJWK (Server operation)
c = g * C # cJWK
K = s * C # dJWK
To recover dJWK after discarding it, the client generates a third ephemeral
key (eJWK). Using eJWK, the client performs elliptic curve group addition
of eJWK and cJWK, producing xJWK. The client POSTs xJWK to the server.
The server then performs its half of the ECDH key exchange using xJWK and
sJWK, producing yJWK. The server returns yJWK to the client.
The client then performs half of an ECDH key exchange between eJWK and
sJWK, producing zJWK. Subtracing zJWK from yJWK produces dJWK again.
Expressed mathematically (capital = private key):
e = g * E # eJWK
x = c + e # xJWK
y = x * S # yJWK (Server operation)
z = s * E # zJWK
K = y - z # dJWK
To understand this algorithm, let us consider it without the ephemeral eJWK.
The math in this example depicts a standard ECDH.
s = g * S # sJWK (Server advertisement)
c = g * C # cJWK (Client provisioning)
K = s * C # dJWK (Client provisioning)
K = c * S # dJWK (Server recovery)
In the above case, the provisioning step is identical and the recovery step
does not use eJWK. Here, it becomes obvious that the client could simply send
its own public key (cJWK) to the server and receive back dJWK.
This example has a serious problem, however: both the identity of the client
(cJWK) and its secure decryption key (dJWK) are leaked to both the server
and any eavesdroppers. To overcome this problem, we use the ephemeral key
(eJWK) to blind both values.
Let's think about the security of this system.
So long as the client discards its private key, the client cannot recover
dJWK without the Tang server. This is fundamentally the same assumption used
by Diffie-Hellman (and ECDH).
There are thus three avenues of attack which we will consider in turn:
- Man-in-the-Middle
- Compromise the client to gain access to
cJWK - Compromise the server to gain access to
sJWK's private key
In the first case, the eavesdropper in this case sees the client send xJWK
and receive yJWK. Since, these packets are blinded by eJWK, only the party
that can unblind these values is the client itself (since only it has eJWK's
private key). Thus, the MitM attack fails.
In the second case, it is of utmost importance that the client protect cJWK
from prying eyes. This may include device permissions, filesystem permissions,
security frameworks (such as SELinux) or even the use of hardware encryption
such as a TPM. How precisely this is accomplished is an exercise left to the
client implementation.
In the third case, the Tang server must protect the private key for sJWK.
In this implementation, access is controlled by filesystem permissions and
the service's policy. An alternative implementation might use hardware
cryptography (for example, an HSM) to protect the private key.