DSS 2026 / Day 2 / Pearl Hall
Forging PLONK Proofs Against an Unchecked Verifying-Key Cache (gnark, BN254)
Vijaykumar Singh
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:00Shall we? Okay. So,
0:03I'm going to talk about forging Planck's
0:05proof against an unchecked verifying
0:07key.
0:08This is the setup in
0:12So,
0:13this work was done by my coworker at
0:16Open Zeppelin,
0:17Vaslin,
0:19aka Vesco.
0:22And yeah, I'm here to present it. All
0:24right? Okay.
0:25So, what is this
0:27this setup about?
0:29Before I go further, these are the
0:31things I'm going to cover very quickly.
0:33So, I'm going to describe the problem
0:35for you.
0:36Then I'm going to go through the three
0:37findings in the
0:39one is
0:40zero cash, crafted cash, both are
0:43both give you a proof forgery, whereas
0:46the the last one is a degenerate base
0:48point, which is a statement forgery. All
0:50right? And then we'll talk about
0:51disclosure and and conclusions.
0:55Okay. So,
0:56I'm going to spend a bunch of time on
0:582.1 and you know, rest we see if we get
1:01time.
1:03All right. So, what is happening here?
1:07So, let's try to build some context.
1:10So, in BN 254 curve, you basically have
1:13KGZ commitment schemes, right?
1:16And
1:17it's a pairing-based
1:19commitment scheme where
1:22the end verification is essentially
1:25[snorts] checking some pairing
1:27equation turns out to be one or not.
1:30So, essentially, if you think about it,
1:32you can see that pairing check
1:35corresponds to
1:37you have
1:39you have E of A one.
1:42So, let me see what what one is. One is
1:44the image. So,
1:46if you recall what pairing is, pairing
1:48is a map from
1:50G1 cross G2 to, you know, FP12 in case
1:53of eight pairings, which is what BN 254
1:55curve uses.
1:57And
1:59so one is generator for G2 group
2:03and
2:04your X is the the toxic waste which is
2:07produced when you you know you do SRS uh
2:10so the initial setup ceremony.
2:12So in the end you you you check
2:15essentially whether this pairing map
2:17holds
2:18uh sorry this pairing equation holds uh
2:20and
2:22if it holds then then you pass the proof
2:24uh if it doesn't then it you know the
2:26full proof fails. Okay so this is a
2:28setup.
2:29Now uh pairings are bilinear maps so
2:32essentially you can uh what you can do
2:34is you can move this X here you get XB
2:36you have A so AXB
2:39uh
2:39should turn out to be so A + XB should
2:42turn out to be identity. Okay
2:44so if A + XB is identity then then
2:46essentially you have the the the the map
2:49will check out all right.
2:51So it all rests on one element really of
2:54G2 uh which is the X uh the toxic waste.
2:57It is unknown to prover it's unknown to
2:59verifier yeah but they they have both of
3:02them have the commitment to X.
3:04Okay
3:05so
3:06so ideally um
3:08so this is a setup
3:10but
3:12unfortunately
3:14quote unquote unfortunately the grow
3:16verifier
3:17instead of basically checking against X
3:20checks against something else which is
3:22the line
3:23in the in the
3:25so let me just clarify what the line is.
3:29Uh so I'll come to the line a moment but
3:31first of all let's see uh what happens
3:33um
3:34so if you recall from the previous slide
3:37we had this uh
3:39equals to identity
3:41right if you go on the next slide
3:43so pairings are are are are are
3:45non-degenerate but if you have EA comma
3:49zero element
3:51of the G2 yeah zero is the identity if
3:54you're thinking about elliptical group
3:55as as an additive group right so if you
3:58do E of A zero then it will always come
4:00out to be one
4:02No matter what.
4:03So, if you look at this map here,
4:07if I had zero and zero here,
4:10it will turn out to be
4:12it will pass, right? Irrespective of A
4:14and B. Doesn't matter what A and B are.
4:16So, can we have that setup?
4:18All right? So, can we have Q zero and Q
4:20one as basically
4:23the additive identity or or the
4:26so-called point at infinity for the
4:27elliptic curve. That is a problem
4:30So now, so the question is, can we have
4:33that?
4:36And the answer is yes. We can have that
4:40because of the following scenario. So
4:41what happens is if you look at the
4:43pairings,
4:44the way pairings are computed for the
4:46GNARK is that you have E of P comma Q.
4:49The Q is fixed. Yeah, that was from the
4:51toxic waste in in the second part.
4:54So what what what what this
4:58computation of pairing does is that
5:01the it runs a Miller loop.
5:03All right?
5:04So and in the Miller loop, what happens
5:06is
5:08there are already some precomputed lines
5:10which are inside the Miller loop.
5:12Okay? So instead of having using
5:15commitment to X,
5:17yeah, the lines which are already
5:19precomputed are part of your of are part
5:22of your computation. Okay? And this is
5:24where the problem comes in. We'll see
5:25that in a moment.
5:26So Miller loop runs and then, you know,
5:28there's a final exponentiation and then
5:29you get if you get one, you pass. If you
5:32get do not get one, it fails.
5:34So the
5:35so the idea is
5:39the following. So
5:42what
5:45GNARK library does is it says that it's
5:48checking for this pairing
5:50where you have basically A of G two,
5:53which is your toxic waste. But
5:55essentially what it does is
5:57it has a verifier already precomputed.
6:00The lines
6:01precomputed given by the prover.
6:04It does not checks it.
6:06And then
6:07if these lines are not properly
6:09constructed,
6:10then there's the issue. Okay? So let's
6:12look at this
6:14So the idea is that we have
6:17lines, we have precomputed lines and
6:20they are never enforced when they are
6:22used in the proof. All right? So so the
6:25the starting the
6:27the issue is at the
6:29even before the start of the proof. It's
6:31in the setup.
6:33The the lines are precomputed. If the
6:35verifier does not checks it
6:38against it being identity all the time,
6:40you basically have the have the fake
6:41proof passing. So so this is the
6:44first issue.
6:46All right. So that's that.
6:50Okay. So second proof is the following.
6:52Second issue is the following.
6:54Let's go back to our equation again.
6:57We have here
6:59AX so we had A + XB equals to identity.
7:03So earlier the issue was what? The issue
7:05was basically you take
7:07you took X to be commitment X commitment
7:10of X to be identity and
7:12the commitment of one to be identity.
7:15Again, do not confuse it. By here
7:17identity I mean the zero element and
7:18that was a problem.
7:20Now the second case can arise if you
7:23have X as
7:26what you can get still
7:28zero if you had X as
7:30some kind of
7:32minus A + B.
7:34Oh sorry, minus A divide by B. Okay. So
7:36that's the
7:37the source of the second issue
7:40as you can see here.
7:43So the idea is you have a toxic waste.
7:47The commitment here.
7:48So the issue is that the attacker will
7:51what will do is instead of
7:53it will it will take some X prime of its
7:55choice in a way that
7:58eventually you you have you turn out to
8:01be this you have turn out to have this
8:03equation A + X prime B
8:06and then you choose B to be minus A or X
8:08prime and then you you essentially have
8:10what?
8:11Again a proof forgery, right? Okay.
8:14So in both the cases
8:17what was the problem?
8:21So in first case you you you don't want
8:24the zero of the elliptic curve or the
8:27point at the infinity
8:28to be the part of the proof. So you
8:29should check that, right? So
8:38so I'll skip the third one uh but you
8:40can have a look uh
8:42so let's see. So how was the issue
8:44basically uh
8:46so on 7th of
8:48July the issue was
8:50drafted and then disclosed to the um
8:54disclosed to the uh Nark.
8:56Uh it was confirmed and basically you
8:59know
9:00um
9:01and then it was acknowledged but it was
9:03considered to be out of scope of the
9:05trusted model which actually makes sense
9:06because it is before the the prover
9:09verifier game assumes that the the
9:11trusted model is honest on the top
9:14but uh
9:16and then on 28th of August it was
9:19approved for the
9:20public discussion.
9:22Okay, so now why do we call it shadow
9:24boxing?
9:26Well, the idea is that in verifier
9:29you are basically
9:31seeing G two,
9:33right? You're you're looking at X
9:35at the commitment of X
9:36but essentially what hits you is a line,
9:39right? Because you have basically these
9:41computation of lines which are which
9:43could be malicious, right? So so so we
9:46do the validation but against the wrong
9:48object. Yeah. Uh yeah, so that's why the
9:51name is called you know shadow boxing
9:53because you are you're fighting the
9:54wrong opponent. Yeah, the the opponent
9:56actually is a line, right? So this is
9:58the the whole setup of the uh
10:01of the
10:03of the shadow boxing um and
10:07yep, I I think you know what?
10:09I pretty good. I finished earlier.
10:12Thanks, thanks for your attention.
10:15>> [applause]
Source and accuracy details
- Auto-generated YouTube captions; native source language tag preserved; not human-verified against audio.
- Names, organizations, technical terms, code and numbers may be misrecognized.
- Source wording and disfluencies are preserved; no editorial corrections were applied.
- Export contains whole-second cue starts only; cue ends and subsecond timing are unavailable.
- A full caption-track export does not establish that every spoken word is represented.
Source export SHA256: 98ca1532f8a16b6d89cf4e2ed2b7e79e402a94cc867e48367b7d6a2555f6b8a8