DSS

DSS 2026 / Day 2 / Pearl Hall

The Attestation Gap: Securing Private AI and TEE-Backed DeFi

Rahul Saxena

14:45 · Recording

Watch on YouTube ↗Open recording viewerOriginal programme entry
Download plain textDownload timestamped JSON

Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.

Transcript

0:02Hello.

0:03Hey everyone.

0:04So it was a great follow-up talk to

0:08start speaking at because

0:11privacy is the need of the hour and for

0:13privacy you can have a bunch of PETs,

0:17privacy enhancing technologies and the

0:19world is a efficient place. It will find

0:22the best solution

0:23between ZK, FHE, MPC, TEEs and what not.

0:28So

0:30let's talk about TEEs today.

0:31I'm Rahul Saxena.

0:33I'm the founder of Bluethroat Labs and

0:35we kind of specialize in TEE security

0:38amongst other things.

0:40So let's dive right in.

0:43The attestation gap, securing private AI

0:45and TEE backed DeFi.

0:47So this talk is necessary because I

0:52I feel that there is a huge

0:55education/awareness

0:58gap in the market around what TEEs offer

1:01and when people say that we use TEEs

1:04that they simply assume that it's

1:07completely private. Something is running

1:10in a computer where the computer itself

1:11doesn't know what is going on. So it's

1:14private and it can be inferred on chain

1:17as an oracle. So you know a complete

1:20privacy solution is up for grabs.

1:23But there is a huge amount of infra that

1:28needs to go to make this happen. There

1:31is a huge amount of steps that need to

1:34be taken to make sure that TEE usage is

1:36actually safe or not.

1:39And that is the attestation gap. Company

1:42A saying that they use TEEs and use

1:44attestations for privacy

1:46versus company B saying that they use

1:48TEEs for attestations

1:50for privacy

1:52can be two completely different claims

1:54and let's now figure out why.

1:56And right now private AI and TEE backed

1:59DeFi is

2:01privacy focused DeFi are the two main

2:03topics where T's are being used very

2:05heavily.

2:07So

2:09when you say that you are using T's,

2:11what are your users trusting?

2:13For private AI you are assuming that all

2:16the data that you are sending to the

2:18models, the inference that is private,

2:20right?

2:21And for T backed DeFi, what are your

2:24users assuming? For example, you cannot

2:26be copy traded. Like let's say you run a

2:29matching engine inside of a TE and

2:32settle things on chain on Hyperliquid,

2:34you get a

2:35private uh market uh

2:38private trading place in Hyperliquid.

2:41So

2:44we will now

2:45take a look at the bunch of research

2:47that we have done in Blue Throat Labs

2:50for writing our TE research T security

2:52handbook and figure out what exactly can

2:54go wrong when you're using T's.

2:57So the first thing,

2:59the attestation gap, the gap between

3:01what attestation proves and what your

3:03protocol promises.

3:04So

3:05when you get an attestation, that is

3:08simply a promise about the environment.

3:11Something ran in a TE,

3:14something was executed and you get a

3:18attestation certificate of that. That is

3:20it. That does not imply anything less,

3:23anything more. Your policy on how to

3:26handle that is the main blocker for how

3:29secure your applications are going to be

3:31because after your policy you do take a

3:34decision on who to give access, who to

3:36give money, who to send information,

3:38etc. etc. etc. So your software owns the

3:41decision to trust.

3:43The evidence, the attestation is simply

3:45an artifact on which you decide you want

3:48to trust further or not.

3:51So

3:52let's see. This is one of the first few

3:55bugs that we found in uh

3:58Fala/Dstack.

3:59So they had an offering for

4:02private AI, and of course in private AI,

4:05what you're assuming is that your

4:07attestations I mean

4:09your

4:10inputs are private. Nobody can read it

4:12and the output you get, nobody is

4:15reading them.

4:16So the software, the logic for that runs

4:20in the CPU and the models themselves run

4:22inside of the GPUs. So both the CPUs and

4:26GPUs need to provide attestations to

4:28prove that

4:29they're running privately. They are

4:30inside of TEs.

4:32And what we found that the CPUs were

4:35secure. They had their attestations.

4:36They were being checked correctly.

4:38But for the GPU report itself, it was

4:41simply getting passed. There was no

4:43checks against the kind of attestation

4:45being received from the GPU.

4:48So without any checks, the users were

4:51getting their back their command. So

4:53technically there was no guarantee that

4:56the prompts that they were giving that

4:58the private AI they were using was

4:59actually private because without the GPU

5:02being

5:04proven to be in a secure enclave, you

5:06can never say that you are running a

5:07private DeFi

5:09private AI.

5:10So passed is not verified. Even if you

5:14get the attestation, the attestation

5:16needs to be checked against certain

5:17fields to make sure that the attestation

5:19is valid.

5:21So again, this was a public finding was

5:25remediated. Falla Security thanked us.

5:27Great.

5:29So this was another finding that we had

5:32in Falla again. So this was the library

5:35that they use called Dcap QVL.

5:38And

5:39here again, what happened was

5:42they were fetching a collateral from the

5:45Intel signatory

5:46and they did not check a bunch of fields

5:50that need to be checked. For like

5:51basically the

5:53coating enclave identity was not being

5:55matched. So even though the collateral

5:58was being fetched, even though a lot of

6:00data from the collateral

6:02I mean a lot of data from the

6:03attestation was being checked. It is the

6:05collateral.

6:07The main thing

6:08who has who is the quoting enclave

6:10identity, that was not being checked. So

6:12there are fields like MR signer, ISV

6:15ECN, etc. etc. that was not being

6:17checked. So again, even when you get the

6:20collateral and you are doing the checks,

6:22if you do insufficient checks, you do

6:25not even know which kind of T or even if

6:27it is a T

6:28that is giving you attestation

6:29certificates.

6:31So partial verification can still return

6:33valid and you can still be in a facade

6:35that

6:36your stuff is running inside a T, it is

6:39secure, it is private, but it is not.

6:41Again, a public finding, we got a CV for

6:43this.

6:44Thanks to part of the security team.

6:47Next.

6:49So this is a very interesting thing that

6:51I want to speak directly to the protocol

6:53owners in the room and

6:55on YouTube directly.

6:57So

6:59this is a KMS and app applied

7:02differential TCB gates. So what I want

7:04to focus on here is that

7:07when you think about TC when you think

7:09about

7:11T's, it's all about

7:14how many layers of security do you want

7:17to get it to? How much paranoia do you

7:19want to give into? And I think

7:21technically that should be defined by

7:23the kind of

7:25TVL that you're supporting or the kind

7:27of

7:29of service that you're offering.

7:30But

7:32there comes a time where you have to

7:34just

7:35stop digging because like it it it comes

7:37to a place like you just cannot

7:41simply just do everything by yourself.

7:43You have to trust at some places.

7:45But it's very important that you figure

7:48everything out like what are all the

7:51things that can go wrong? What are all

7:53the threat

7:54threat actors if they lie, if they

7:56behave maliciously,

7:58our protocol can fail.

8:00For example, here

8:02there was there's something called TCB,

8:04the trusted computing base, and Intel

8:07can sign a

8:08certificate, it can give an attestation,

8:10but also give your status that the TCB

8:13for this attestation is out of date.

8:15So the trusted computing base from which

8:17this attestation was generated was out

8:19of date. So it is now up to you whether

8:22you want to process this attestation,

8:25pass it through your status check, your

8:27policy and take actions or not or or not

8:30accept it.

8:31So

8:33why I want to bring this up is that

8:36in the ideal world you will be doing

8:38everything by yourself, you will be

8:39setting up your own infra, you will be

8:40setting up your own

8:42policies and whatnot, but in the

8:45beginning in the real world you would

8:47you will depend on some

8:49TE provider

8:51and then you have to figure out all

8:53these checks whether they are being done

8:54by the provider itself or they are being

8:57left to the app developer as your own

9:00policy advice as your own policy

9:04choice.

9:05So you have to figure out what are the

9:06decisions that are in your hand.

9:09Take it and do not assume that the TE

9:12provider is doing it for you.

9:14So again, this was not

9:16a bug per se, but

9:19the

9:20team at Dstack and Falu, they took care

9:22of it and

9:24it made it more secure.

9:28Now we come to the next level of TE

9:31bugs.

9:32So just because something is running

9:34inside of a TE does not mean that the

9:37software inside it is bulletproof.

9:39If the software inside the TE itself is

9:42buggy and those bugs can be triggered,

9:45then the TE does not protect you.

9:47For example, this is one of the bugs

9:50that we found in an audit of Secret

9:52Network. Secret Network is an L one and

9:55it uses Intel SGX TEs in its nodes.

9:59[snorts] So [clears throat]

10:02an active validator during its proposal

10:04slot can use that particular byte, just

10:080 x FF, that single value

10:10and its proposal slot send propagate it

10:13during its proposal slot and then when

10:16all the honest validators try to decode

10:19it

10:20they would hit a panic and then they hit

10:21a panic they cannot reach final T

10:23leading to a chain halt.

10:25The T attestation were completely fine

10:28that attestations were being checked the

10:30policies were correctly implemented the

10:32trust model was completely fine but the

10:34code itself was not able to handle

10:37malicious input so this is what

10:39happened.

10:41The enclave passed bytes before normal

10:43decoding so in in in the time of the

10:46proposal

10:47the active validator set one of the

10:51Byzantine validators

10:52can put the 0x ff value there

10:56it can reach the consensus and the

10:58consensus software typically does not

11:00care about the payload that you have it

11:02wants to take care of

11:04the consensus guarantees consensus

11:05property so it does not read it

11:07and it forwards it to the SGX parser

11:09directly SGX parsers

11:12tries to parse 0x ff

11:15protobuf cannot figure out what to do

11:17and it panics which leads to

11:20a chain halt.

11:22So this was a chain halt bug that we

11:24reported in secret network and this is

11:26all just to point out that we use T's do

11:30not make you secure T's is just a part

11:33of the equation and there is a lot of

11:35surrounding work that you have to do to

11:37make it actually secure.

11:41Yeah and also this is a very cool

11:44way to tell you that

11:47the boundaries of T's and some other

11:50systems is a place where a lot of bugs

11:53are found typically

11:55because

11:56any boundary should be treated as an

11:58oracle no matter how trusted that

11:59boundary is and if you do not trust it

12:02I mean if you trust it and you just you

12:06let your T's get any inputs you kind of

12:08get end up getting wrecked.

12:13Cool so

12:14we had a lot of findings. You can

12:17find them at bluetoothlabs.com/releases.

12:20But from these four main findings, what

12:22I can tell you are the four checks for

12:25your T [snorts] integration in your

12:26protocol. Okay?

12:28So,

12:29it's pretty simple. Alter or remove

12:31required evidence, which is rejected

12:34invalid or missing evidence.

12:36Basically, if the evidence coming in is

12:38not actually coming in, it's not

12:41correctly coming in, it's not

12:43fresh, it's not by the correct entity,

12:46reject it.

12:48Present a disallowed entity or status.

12:50Now, check the status of the signature

12:53itself. Like the attestation might be

12:54valid, but what are the values inside of

12:57the attestation? What is the collateral

12:59in the attestation? Check all of them.

13:01If they are correct, let it through.

13:04And it's not always going to be binary.

13:06You have to make decisions based on your

13:08protocol requirements, whether you want

13:10to

13:11let a certain attestation do a certain

13:14state change in your protocol or not.

13:16So, again, do it.

13:18Third is no acceptable evidence

13:20available, [snorts] which is

13:22if nothing is present that authorizes

13:25something from a T, do not process it.

13:27And

13:29send hostile enclave input, which is

13:32always be mindful of all the inputs that

13:34are coming into your T's from any

13:36boundaries whatsoever. Nothing is

13:38trusted for a T.

13:40So,

13:41apart from this, like for the T itself,

13:43there are a bunch of other things. For

13:45example,

13:47how do you make sure that the code that

13:49you audited is the actual code running

13:50inside of a T? How do you make sure that

13:53the

13:54clock that the T software is is relying

13:57on is actually correct because the time

14:00can be supplied by the hypervisor, which

14:02is not trusted in the T model, etc.,

14:04etc. But overall, in a general sense,

14:06for T integrations, this is a pretty

14:09cool checklist.

14:13And finally, I would say that we have

14:15written a really cool T security

14:17handbook,

14:18which covers all the bugs that we have

14:21found, all the knowledge that we have

14:23gained around T's. You can scan this QR

14:26code if you want to read it. This is the

14:29URL and we have also applied for

14:32a grant and in the DAO fund

14:35Ethereum funding. So if you like the

14:37book, please donate anything that you

14:40can.

14:41Thank you. Have a nice day.

Source and accuracy details

Source export SHA256: c4424621b3334d9b43bb2dc7df9daa14ad84f325db34f7f33ae2f878bc446544