▲ 141 points
back
63 comments
"At first glance, this seems fine, but it is not. If an attacker knows the plaintext p1 and the ciphertext c1, then they can compute the keystream by XORing p1 and c1 together"
Also, if the attacker only has c1 and c2, if the nonce is reused then c1 xor c2 will be the same as p1 xor p2. In most cases, two plaintexts xored with each other are trivial to decode.
I made an entire "game" based on this concept (AES CTR though) https://aes-cpa.fly.dev/
For anyone interested, I made a few more:
* https://ctr.var.tailcall.net/
* https://ecb.var.tailcall.net/
* https://cbc.var.tailcall.net/
The goal is a bit different (I use them when teaching university courses, to show that encryption is not authentication), but the same ideas apply.
Not just AES-GCM.
The vast majority of encryption algorithms must be used in a nonce-respecting scenario. This is part of the contract to achieve the claimed security properties.
Alternatives require multiple passes over the data, which is not applicable to some protocols in addition to having performance implications.
Common protocols such as TLS transparently handle nonces in a safe way. But the primitives used in TLS may require additional steps to be safely used is other contexts, especially in distributed systems.
Whenever applications try to use these primitives directly, using a fixed key and picking nonces at random is a very common practice. Unfortunately, due to their small size, nonces collisions can quickly happen.
We're missing standard constructions with large nonces that would alleviate this problem, because IETF protocols haven't needed them. But there's a lot of evidence that many custom applications and protocols do.
There are multiple great proposals to derive AES-GCM subkeys and nonces from a key and a large nonce. We may expect convergence and adoption in crypto libraries soon.
Until then, constructions such as XSalsa20 and XChaCha20 are widely implemented and deployed. If you don't need NIST compliance, they're excellent choices.
But my recommendation today would be to replace AES-GCM with the AEGIS family of algorithms whenever possible. They have nice properties that AES-GCM doesn't have, including more comfortable usage limits, much better performance and large nonces up to 256 bits.
This page [1] and that draft [2] summarize usage limits of common constructions, including when using random nonces.
[2] https://doc.libsodium.org/secret-key_cryptography/aead
[3] https://datatracker.ietf.org/doc/draft-irtf-cfrg-aead-limits...
I find the article a little confusing. IMHO the point is that you must NEVER reuse an XOR key sequence for stream cipher encryption. With RC4, this meant that you could never use the same key. With modern stream ciphers there is the nonce for this - the CTR mode of a block cipher is also a stream cipher. (GCM mode is just an extension of CTR mode for authentication).
I've put together a little online demo tutorial (in my teaching and learning programming language).
GCM's nonce reuse failure modes are worse than CTR's.
I dunno, nonce means number-used-once, it should be kinda obvious that it should be used only once?
Correct. However, some implementations actually incorrectly refer to the nonce as an "IV" (initialization vector), where it's not so obvious.
Also, it's not entirely clear just how bad a reuse actually is. For example, in AES-CBC, reusing the IV has much less impact than reusing the nonce with AES-GCM.
NIST calls it an IV (or at least did when it came out).
And yet...
Aside from just not understanding it, it's plausible that someone would generate nonces weakly, say, from a weak source of randomness.
Even using a strong source of randomness for an AES-GCM nonce is weak over enough messages, since it only gets you 48 bits of collision resistance.
If you're not using random nonces, maybe you want to use a counter, and then you have to worry about race conditions, state resets, etc. (if your system lost power immediately after using nonce n, would it boot back up and reuse it?)
> Aside from just not understanding it, it's plausible that someone would generate nonces weakly, say, from a weak source of randomness.
This is actually generally fine for nonces (used in CTR and GCM modes, and in ChaCha20). Typically the only requirement for a nonce is that it is only used once. It is even safe to use a simple incrementing counter.
IVs, on the other hand, are required to be cryptographically random.
My point is that a random AES-GCM nonce is a problem because you will end up re-using nonces by chance, assuming multiple messages are being encrypted with the same key. Due to the small nonce size, this is still a problem even with secure randomness, it's just even more of a problem with weak randomness.
The consequences that using the number used once twice entail differ wildly between cryptosystems though.
I've seen the problem surface in designs where a single key is shared across many devices.
Also depending on the cryptosystem, nonce requirements vary wildly. For example using 1, 2, 3, 4... as a nonce for CTR is a recipe for disaster, but it's fine for many asymmetric ciphers.
I'm curious for the use-cases where people have to maintain a key over a long period where the choice of nonce can't be made strictly non-decreasing or otherwise prevent nonce reuse (per key).
I can imagine VPNs or other packetized communications potentially running into this problem, e.g. with N parties needing to encrypt messages under the same key to each other without coordination on nonces. The worst case I can think of is a large number of devices with a baked-in key and secure RNG but no non-volatile storage. They can't generate more than 2^48 messages with AES-GCM or risk collision.
Full disk encryption has always had a similar problem; generally a single long-lived master key that individual sectors or blocks are encrypted by, often without the additional storage set aside for IVs or nonces (which would break exact sector to sector mapping of encrypted virtual disk to plaintext disk). That leaves IV-derivation to be static per block offset/number, or key derivation on master key and block offset/number.
Devices without secure RNGs are also at risk (microcontrollers with no non-volatile storage that restart a lot, for example).
I'm curious if there are any other hard cases where nonce reuse becomes a risk in practice.
Great post! Thanks for taking the time to put this up.
What do you think the ratios are regarding improper use of nonce with this mode?
Most implementations that I am familiar with intentionally generate a random nonce to help lower the percentage of app devs doing this very thing
My company need deterministic encryption to search encrypted data.
Turns out the people who wrote the in house Go library didn't have any idea. There is no non-deterministic encryption function because that might be too complicated for non-senior engineers (afterall they wrote most of the actual application) to correctly choose.
The first version use AES-CFB. There's no authentication. It's probably copy pasted from a public Gist and nobody ever commented on it that it is insecure. I wonder if it was actually intended to be the non-deterministic version, but the higher level wrappers do not wrap this function so people didn't actually use it.
The second version use AES-GCM with nonce derived from the key and AD. Since nobody understand why AD is needed, AD is always nil. Essentially there's ever one nonce.
I think the problem is that many senior engineers know that encryption use "AES" library but the Go standard library doesn't tell you how to use it securely.
Surprisingly this mistake also happen in our Java stack that was written by a different team. A senior engineer did notice and quietly moved away from the vulnerable version without telling the Go version.
I wrote a POC to decrypt data of the Go version, then wrote the third version, perhaps it will be open source soon. The new library only implement envelope key management, encrypted string wrapper and ORM integration. The rest is Google's Tink.
> My company need deterministic encryption to search encrypted data.
I'll take things you should never do as a non-expert for $100.
> The first version use AES-CFB. There's no authentication. It's probably copy pasted from a public Gist and nobody ever commented on it that it is insecure. I wonder if it was actually intended to be the non-deterministic version, but the higher level wrappers do not wrap this function so people didn't actually use it.
Lack of authentication is probably the least of your concerns if your product is searching over encrypted data.
You use the AD to authenticate additional information that doesn't need to be encrypted. For example, if you separately encrypted every record of a database, you could leave a non-sensitive identifier exposed along with each of them and validate it as the AD when decrypting. This would allow you to find specific records quickly assuming you also had an (encrypted) index or some prior knowledge. As with any case of leaving some data exposed, this can open up certain avenues of attack depending on the threat model. If the data can be tampered with, for example, this isn't a good idea since an attacker can corrupt your database (you'll know, but it will be unusable).
[Edit: I was unaware of the existence of "deterministic AEAD" before I wrote this: "Deterministic" encryption is discouraged because it passes through block-aligned patterns in the plaintext to the ciphertext. There is a simple method to do what you're after: it's just feeding your data (with padding) directly into the cipher (so-called ECB mode). Go's standard library gives you the raw AES cipher to do this with, but it doesn't expose the standard padding mechanisms (and it's not authenticated). You should be aware that doing anything like this leaves your data open to certain kinds of cryptanalysis that can infer the plaintext without directly breaking the cipher.]
I largely agree that the standard library doesn't provide any solid guidance or higher-level APIs for any use case other than TLS. The implementations seem to be pretty high-quality but you quickly go from "it's hard to use this wrong" in some libraries to "here's a drawer full of sharp knives" in others.
Correct. GCM is an improvement over ECB and CBC; it doesn't magically transform a symmetric algorithm into an asymmetric one. So most libraries are going to focus on the use cases where symmetric crypto makes sense, which are single-party scenarios such as disk storage. Google's Tink library, for example, completely hides the nonce parameter from its API.
GCM is an improvement over CBC since it has authentication, but it does have a few weaknesses that CBC does not suffer from:
1. CBC does not have the same class of vulnerability to Nonce/IV reuse. Reusing an IV would leak some information about the first block (or first few blocks which are the same), but it would not give your a XOR of two plaintext or let you recover the keystream. On the other hand, CBC is vulnerable when IVs are predictable (e.g. the BEAST attack).
2. CBC with a proper encrypt-then-MAC scheme (e.g. HMAC-SHA256 + HKDF-SHA256 for generating Authentication and Encryption Keys) can encrypt more data than GCM without rotating a key. GCM with random nonces are particularly problematic, since at one point you would run into a nonce collision.
Overall, AES-GCM is preferable to AES-CBC because it is quite hard to implement a good encrypt-then-MAC scheme on top of AES-CBC unless you know what you're doing. But it's not good enough as a general worry-free solution, even when you're using a library to wrap nonce generation for you. What you want is XChaCha20Poly1305, if you're going for an ubiquitous and mature cipher.
Fair points. AES_GCM_SIV [1] is my choice where it's supported, personally, which is nonce-reuse-safe (wrt to key material leaks). Plus at least the primitive is hardware-accrlerated more often than not.
The problem with a random nonce is that most implementations also use a nonce of 12 bytes which under some use cases might not be enough before you repeat a nonce. So to remedy this they suggest using a counter but this could be hard to implement.
When I use AES-GCM I just use a bigger nonce and use a random one.
Last time I used AES-GCM I had a really hard time getting the person writing the other end to not re-use nonces.
> nonce and use a random one
Recently discussed: "Galois/Counter Mode and random nonces" (28.05.2024) https://news.ycombinator.com/item?id=40497525
> When I use AES-GCM I just use a bigger nonce and use a random one.
I don't think nonces bigger than 12 bytes will help. My quick reading of the AES-GCM spec is that when using a nonce that's not 96 bits (12 bytes), it is hashed to 96 bits. So either the nonce (called iv in the spec) is carefully constructed from a counter and set to exactly 96 bits, or the number of invocations is limited. The spec still restricts use of a key to 2^32 total uses for random nonces of any bigger length (resulting in a re-use probability of about 1e-10):
https://nvlpubs.nist.gov/nistpubs/Legacy/SP/nistspecialpubli...
I think the writer is @frereit, they submitted 2 days ago https://news.ycombinator.com/item?id=40623885
Yes, I am, but unfortunately I do not think I can provide any answers here. A quick internet search reveals some CVEs for nonce reuse.
If I had to, based on absolutely nothing but a gut feeling, guess, I'd think this may appear more frequently in IoT devices, where AES-GCM is attractive because of its speed, but randomness is sometimes in low supply?
It's worth mentioning AES-GCM-SIV[1], which is the fix for this issue.
The alternative, which I prefer, is an XGCM-like construction that just gives you a large enough nonce to comfortably use random nonces.
+1, soatok has a write-up of how that works: https://soatok.blog/2022/12/21/extending-the-aes-gcm-nonce-w...
...a variant on that is DNDK-GCM in draft at https://datatracker.ietf.org/doc/draft-gueron-cfrg-dndkgcm/ and a recent presentation: https://youtu.be/GsFO4ZQlYS8 (this is Shay Gueron who worked on AES-GCM-SIV too).
AES-GCM has a 12 byte nonce if I recall correctly. Is 96 bits of entropy insufficient to guarantee uniqueness every time it’s generated?
Could this be extended to give us XOCB? I am not sure it would make much sense with the OCB size recommendations.
The "fix" is to use a nonce misuse resistant cipher, of which AES-GCM-SIV is one.
But, AES-GCM-SIV requires two passes over the data, which isn't always ideal.
The goal of the CAESAR competition [1] was essentially to find alternatives. Whether that goal has been met is a bit unclear at the moment.
> The goal of the CAESAR competition [1]
At this point OCB has an expired patent, and only needs one pass over the data:
From the OCB FAQ[1]:
>What happens if you repeat the nonce? You’re going to mess up authenticity for all future messages, and you’re going to mess up privacy for the messages that use the repeated nonce.
The loss of privacy on OCB nonce reuse is not as severe. It would be more or less the same as with ECB mode.
GCM is well known to have sudden death on nonce reuse, which is why we used GCM-SIV for the ZeroTier v1 protocol. (Our new protocol that stiiiil is not in product is Noise based.)
Nonce reuse in SIV is much less catastrophic. Reuse of a nonce can reveal if two packets are identical but doesn’t let you do anything else.
Of course there are categorically better modes than GCM but they are not widely supported. ChaChaPoly is better cryptographically but no hardware acceleration, which matters on small devices.
If the attacker never gets a hold of a plaintext-ciphertext pair, how well does AES-GCM with nonce reuse hold up?
It breaks down to a primitive of repeating key xor.
If you never had a plaintext you could potentially (depending on the content) collect enough ciphertexts to do frequency analysis on it. You'd recover the keystream at least partially, and then guess based on context to fill out the rest.
I understand that nonce reuse is catastrophic, but I don't think I understand when it can be abused. Does the attacker have to know which two messages share a nonce? Is knowing that out of N messages, at least one pair shares a nonce already enough?
Well, the nonce is (usually) public information. It is shared along with the ciphertext, so that the other party can use the same nonce to validate and decrypt the ciphertext.
So it is trivial to detect which two messages share a nonce, if any do.
>T1 ⊕ T2 = ((U10 ⨂ H3) ⊕ (U11 ⨂ H2) ⊕ (U12 ⨂ H) ⊕ Ek(y0)) ⊕ ((U20 ⨂ H3) ⊕ (U21 ⨂ H2) ⊕ (U22 ⨂ H) ⊕ Ek(y0))
= ((U10 ⊕ U20) ⨂ H4) ⊕ ((U11 ⊕ U21) ⨂ H2) ⊕ ((U12 ⊕ U22) ⨂ H).
Shouldn't the result be ((U10 ⊕ U20) ⨂ H3) ⊕ ((U11 ⊕ U21) ⨂ H2) ⊕ ((U12 ⊕ U22) ⨂ H) ?
> I don't think I understand when it can be abused
The same key + nonce generates the same keystream.
The ciphertext is generated by xoring the plaintext with the keystream.
The keystream can be recovered by xoring the ciphertext with the plain text.
To abuse it...
The defender needs to re-use both the same key and nonce.
The attacker needs to have a ciphertext/plaintext pair, know or find the position of that text in the keystream, and needs access to other ciphertexts generated with the same key/nonce.
nonce really should be renamed. if only for british slang avoidance.
I agree on principle, but I feel the next name will also be turned into another British slang term for sex offenders, if only because some lad will find it funny and become determined to see it through.
We are, after all, talking about the country that terrorized an Austrian town into changing its name after decades of sign theft and jokes [0]
It's also the language/dialect where you can basically used any phrase ending in -ed to mean "drunk". The classics being pissed, plastered, wankered. But if someone came up to me and said "I got (wardrobed|hadron-collidered) last night" I'd know exactly what they meant
Excellent article was a very insightful read.
alright, whomever owns this website got me with that "Activate Windows" footer.