WTF is a TEE? The Room That Locks From the Inside: TEEs Explained

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:01

   For the past decade, data security has really been about encrypting data in transit and data at rest. But with AI, we have a new kind of problem. Data can't be encrypted when it's actually being processed by AI. So, we need a new type of system to do that. System's been around for a while and it's become a reality now for companies who actually leverage this technology is called TEES or trusted execution environments. This

2. 0:25

   is where the operator of the actual machine is locked out. That means is that even those who are trying to get into the system still can't see it either. So it's protecting from both sides of it. But it allows the data to actually be processed that is usable for a system like AI because again it cannot be encrypted when being processed. Let's go over how this works. Your data goes into a room that locks from the inside.

3. 0:49

   And that's basically the analogy that you want to understand here of how a trusted execution environment works. A question goes into what's called a sealed processor. And from there uh your answer will then come out of that by locking out the hardware. The company who is renting you the computer can't see it. Uh the administrator with root privileges machine still can't see that information. Anyone who walks into the

4. 1:14

   data center and picks up that server still can't see what's happening on there. Uh it doesn't even the people who built the service has no way to see what is being processed in there. And then on top of all of that, that hardware actually provides you a receipt that you can check yourself, which I will show you on this video I like to use here is it's like a soundproof booth. Basically, you can slide information into that booth. They can process it and give you information back. And what happens inside that booth

5. 1:41

   is unknown to any party. Importantly, at the end of it, you also then have proof of what actually happened. As I mentioned at the beginning of the video, this is critical because encryption has a hole in it. when you consider how it would work with as I mentioned at the beginning of the video encryption has a hole in it and this is something that is important for AI processing. So here you can see the model of data can be at rest data in transit and then when it goes to

6. 2:05

   the model it has to be in plain text. It can't be unreadable anymore at that point. The model has to be able to see it. uh partly if you were to have a system that's fully encrypted from end to end and even when it's being processed the overhead is just too big the processing power is too big to actually have any kind of meaningful results. So this is the way that you're able to actually execute on this.

7. 2:26

   Difference with the TE now is that in a traditional setup where it's just a database reading, it doesn't quite make sense all the time because it's constantly having to do transactions and the back and forth all the way through still would have latency using that trust execution environment. But with AI because you are sending it kind of a single verification, it's processing.

8. 2:50

   it's only paying once as it's doing that. So tax on the overhead of using a large model inference is then minimized. Another way to look at it is that the box is opaque from both sides. So in that trusted execution environment, it is sealed so that no one can ever see what's inside there. Even if you own the machine, you still cannot see what's in there. So both your words and the actual the weights of the model, all of that

9. 3:14

   stuff is all hidden from this. So, how do you actually go and check this and how do you actually prove that this is happening when you make a call to someone who claims that you're using TE? Now, the ability to actually make that call, it has to be opened up by the instance. And so, uh, if they were blocking you from making that call, that is one that would be hard to actually prove, but that'd be silly for them to actually do that because they want you to see that this is actually viable.

10. 3:42

    But, this is how this would work. So, you're checking a promise you were never asked to believe. And so uh basically what you do is you have a randomized number and it goes to that that sealed chip. Uh it measures itself. It then signs that transaction. The signed report comes back and again it should have that number that you gave it. You can then go again to whoever made that silicon so that you know so it's in Intel or Nvidia. Uh is the signature

11. 4:06

    real? So proof of that it is genuine hardware. So now we just have the actual question. Anything else that you send is going to be fully private. So let me show you how you can do this and I'm going to do this with neari that uses tees for their instance. So in this setup uh I am going to check on their deepseek model to see if it is actually running on a te and I am going to call that endpoint for the addestation report

12. 4:31

    and it's going to do what I just showed on those steps. [snorts] From here you can see that there is producing uh evidence of the different endorsements and certifications. The signing key is generated inside the enclave. So it's not exported. So we shouldn't see that which is great. Uh we can see okay you know which the checks mark off. Does it actually commit to the signing address? Does it commit to our nons? So it actually fresh. It's not just giving us some cached response. So well we can see the architecture here.

13. 5:05

    Everything all the way through. For a long time, trusted execution environments have really just been talked about as a research topic. It has been something that has not been widely adopted in most computing in 2026. Now though, we have seen this actually come about from a lot of different companies.

14. 5:23

    So not just the smaller ones who are trying to do private AI, but you can see even all the way through uh Meta introduced this back in May of 2026 with their incognito chat on WhatsApp. They are running through enclaves. Uh Apple has announced this at the recent developer conference talking about private cloud computing. Uh they are leveraging again Intel TDXs and Nvidia Blackwells in order to do this as well as Google Titans. Uh so they are really

15. 5:49

    leveraging all this stuff for that private executed uh computing. Google cloud has been doing this as well. Uh they announced recently that confidential GPUs will go into their preview model. So now eventually uh those will eventually go into full production but right now they are testing that and opening that up as well. And if you want to build with this today a companies like nearai already have this available so that you can

16. 6:13

    actually like I just showed you can do your verification from there and all of their your inference is happening in that trusted execution environment. What's great about this is that privacy has stopped becoming a policy and it's really become an endorsement. uh these major companies all leveraging this as well is an important step in having better data security and better privacy for the information that you're trying to pass along to your AI system.

17. 6:37

    Hopefully from this video you've learned what a TE is and what trust execution environment actually provides you and how it works. If this is something of interest to you, I would suggest you check out near.ai and see how you can actually leverage TEES in your deployment today.