Hacker News Reader

Comment by mehrdadn | Hacker News Reader

I feel dumb for asking this, but is this really something that's not supposed to be possible from software? My assumption was always that if you overlock, undervolt, etc. then (a) you already have high-privileged code running, and (b) anything can go wrong from crashes to physically breaking your CPU, so I'm not really shocked (not sure if pun intended) to hear undervolting can damage the software state. Should I be?

View on HN

It's not a dumb question, not at all.

Processor vendors have been trying to tell us that they can protect parts of the CPU from its user. Intels version of that is called SGX.

There are very few good use cases of this. Also it doesn't really work.

But there are gazillion ways to attack it, so plenty of papers can be written about it.

To add to this ^ imagine that your enclave is computing a wrong result and then signing this result along with an attestation that it ran the correct code.

This could be fine if the computation first goes through a consensus mechanism that tolerates faults, but could be devastating otherwise.

Sounds like a security Perpetual Motion Machine.

> you already have high-privileged code running

I just glanced at the abstract, but they seem to be attacking the SGX enclave, which is supposed to be isolated even from privileged code.

...in other words, they attacked the horribly user-hostile anti-feature whose real practical purpose is DRM, and so I certainly think nothing of value was lost.

https://www.gnu.org/philosophy/right-to-read.en.html

There’s been a couple of non-DRM uses of SGX, which I’m sure you’ve seen. And regardless of what it’s being used for, I think you know it’d be more productive to talk about the attack rather than potential uses of what they’re attacking.

Is SGX actually used by DRMs? I read that TPMs were designed for DRMs as well and ended up being used for whole different purposes (secure boot, disk encryption)

I read somewhere that Netflix drm runs in sgx and they're basically the only ones.

I read that Netflix doesn't use SGX.

There are many interesting uses for enclaves, most of which have nothing to do with DRM. A good example is Signal’s secure value recovery mechanism: https://signal.org/blog/secure-value-recovery/

Really interesting use of SGX. It says at the time of writing that they haven't decided if they'll use it in Signal yet, are there updates?

Not that I know, I suspect a limiting factor is that there aren’t many cloud providers with good support for SGX at the moment.

It is interesting and useful as a workaround for existing situation, ie running under untrusted OS. It’s not useful as a proper solution (as opposed to a hack/workaround). Except for DRM.

Why not? An OS is a large beast, it’s a massive TCB. Much easier to audit a small amount of code running in an enclave with few dependencies.

Because, as shown by SGX, it doesn’t work. And if it did work, it would lead to a whole new magnitude of the malware problem - malware protected from you by hardware.

The right way to do it is the other way round: have a trusted hypervisor and run your untrusted OS in it. See TrustZone for example.

I’m not sure why you think that SGX shows hardware enclaves “don’t work”. I also don’t see why you think enclaves “protect the malware from you”. Enclaves are created and started from host code, which can interrupt or terminate them at any time.

The scheme you suggest, which isn’t typically how TrustZone is used, gives zero integrity and confidentiality guarantees for applications. I don’t know if it’s “the right way” for some threat model, but for the most typical TEE use cases which are trying to establish strong integrity and confidentiality guarantees in the presence of an untrusted host, it’s absolutely not right nor useful.

Critical flaws in SGX are being found all the time. It's a failed experiment that someone put into commercial products. At the same time there are no examples of SGX-like architecture that actually do work.

Enclaves are started from host code, but host code can't see into them. In other words you have no way of telling whether the enclave you've started is what you wanted, or if it includes malware.

The scheme I've described does give integrity - after all, that's precisely how the actual trusted components work, SecureEnclave, TEE etc: you have a trusted hypervisor, and then run your secure components outside the untrusted OS, in a trusted environment in that trusted hypervisor. It gives you everything SGX could, without its fundamental design problems.

I don’t think it’s helpful to confuse side-channel or micro-architectural attacks with attacks on SGX itself. Stating that hardware enclaves don’t work and do not ship is absurd, they are present in virtually every modern phone for one thing.

Code running in an SGX enclave is measured and absolutely known at enclave launch. The fact that enclave memory is encrypted for confidentiality is unrelated.

I don’t understand why you think trusting the hyper visor is helping anything. You are still open to this attack, and to all side channel attacks as soon as you run any untrusted code.

Which phone uses enclave on top of untrusted OS, as opposed to running the enclave on the side of untrusted OS?

Every Android phone which has an ARM chip with Trustzone.

Which doesn't run on top of on untrusted OS, like with SGX, but rather on the side, which is what I'm describing.

Both SGX and Trustzone are TEEs, which are segregated from the rich execution environment (the untrusted OS)

What you’re saying is a different threat model: your application goes rogue. SGX and TEE in general attempt to solve the reverse: your host goes rogue.

Research has shown that it is not a panacea, but we already knew that. It’s hardware not a full proof cryptographic solution. Some solutions have enclaves gather their results in a fault tolerant way to increase security even more.

So we could say that Intel and hardware vendors in general are looking for a solution that doesn’t exist. Or we can say that this is greatly improving your option when you are really scared of host compromises in your product.

Not sure why you're being downvoted. Software running on the user's machine has no business keeping secrets from the user. This "feature" is the antithesis of computing freedom as we know it.

Trusted computing is at least salvageable. The issue is not the technology but who owns the keys. If users can install their own keys, the technology will empower them with increased security. SGX cannot be used to empower the user like this. It's specifically designed to protect software from the user.

No, the user is still in control of what they execute on the machine, whether it is run in enclave or not. If anything, because it is deliberately unable to patch itself, software running in an enclave gives more control and auditability to a user who can know exactly what code they are running.

Importantly, a user who does not fully trust the machine administrator can still maintain integrity and confidentiality over their computation.

SGX memory encryption keys are ephemeral, they are generated at boot, and they do not need to be owned by anyone to be useful, on the contrary!

Reply on news.ycombinator.com