Are we nearing any sort of consensus that any form of speculation is bad? Is there a fundamentally secure way to do it?
I think for "normal" users the main risk is JavaScript, which can (kind of) be mitigated in software without affecting the rest of the system, so no one really cares about these attacks. But the fundamental abstraction leak between physics and programming will always be there.
https://crypto.stanford.edu/~dabo/papers/ssl-timing.pdf
That's on the local network. I remember a paper doing this over the internet, but couldn't find it. A similar one over the internet:
https://www.usenix.org/conference/usenixsecurity20/presentat...
But in practice it's going to be malicious JS running in your browser: https://www.schneier.com/blog/archives/2021/03/exploiting-sp...
> The GoFetch attack is based on a CPU feature called data memory-dependent prefetcher (DMP), which is present in the latest Apple processors. We reverse-engineered DMPs on Apple m-series CPUs and found that the DMP activates (and attempts to dereference) data loaded from memory that "looks like" a pointer. This explicitly violates a requirement of the constant-time programming paradigm, which forbids mixing data and memory access patterns.
> To exploit the DMP, we craft chosen inputs to cryptographic operations, in a way where pointer-like values only appear if we have correctly guessed some bits of the secret key. We verify these guesses by monitoring whether the DMP performs a dereference through cache-timing analysis. Once we make a correct guess, we proceed to guess the next batch of key bits. Using this approach, we show end-to-end key extraction attacks on popular constant-time implementations of classical (OpenSSL Diffie-Hellman Key Exchange, Go RSA decryption) and post-quantum cryptography (CRYSTALS-Kyber and CRYSTALS-Dilithium).
It sounds like the OpenSSL code is still constant time (it doesn't expect pointers in the intermediate values, to OpenSSL it is just data, not something it will ever dereference) but the attacker-controlled "monitoring" code's runtime changes based on whether the DMP ran or not.
If that's right, then the attacker still needs to run their monitoring code on the target, it isn't sufficient to just run OpenSSL etc itself.
Edit: it is more explicit in the paper, they assume that OpenSSL (the victim) is still constant time:
> In this paper we assume a typical microarchitectural attack scenario, where the victim and attacker have two different processes co-located on the same machine.
> For our cryptographic attacks, we assume the attacker runs unprivileged code and is able to interact with the victim via nominal software interfaces, triggering it to perform private key operations. Next, we assume that the victim is constant-time software that does not exhibit any (known) microarchitectural side-channel leakage. Finally, we assume that the attacker and the victim do not share memory, but that the attacker can monitor any microarchitectural side channels available to it, e.g., cache latency.
But I think you might overall be right that this requires two colocated processes: the paper talks about how the DMP breaks assumptions made by "the constant-time programming model", and I took this to mean that constant-time algorithms aren't constant-time any more. Reading more closely, I think maybe the issue is that "the constant-time programming model" was also assumed to make secrets safe from cache timing side-channels leaking the secrets to other processes on the same CPU, and this seems like it might be the assumption that's broken by the DMP...
I'll have to read more, I've just skimmed the abstract and introduction so far.
In any case, the stack of "this could not possibly be workable / but with enough statistics it can, and computers are plenty fast to generate statistically useful case numbers" is truly mindboggling with these attack vectors.
"More data" might be not practical.
Examples (I'm running Debian)
The kernel, the X11 server, terminal, ssh, bash, anything coming from the official Debian repos including the password manager: in the trusted environment.
browsers, electron apps, anything installed from unofficial repos or language and package managers (npm, rvm, asdf, etc): in the untrusted environment.
It reminds me of mainframes and their redundant and compartmentalized hardware architecture.
> X11 server
Those can very easily execute untrusted code.
The risk here is that there are more individuals with the skills to take this type of attack and bring it to a browser near you.
One apps data is another apps code.
Unfortunately, I've had no luck getting others to buy into the idea that they should understand this level of detail so they can make these calls. Quite frustrating and depressing, since companies will relentlessly exploit their indifference.
If the only way to get trackers on the average person is to serve it from the same first-party domain, or to bundle it in with the giant 'app.js', you better believe they'll do that.
Right now, the fact that only a small fraction of people run adblockers, and an even smaller fraction block javascript, is what allows it to work.
In my case when every attempt fails I know it could be the side effect of some other privacy add on. If it's a random blog/news, that's the end of it. If I really have to use that site I open Chrome, do what I have to do, close it. Of course given a choice I pick sites that work well with only JS from a few inevitable sources.
I would like to salute my shitty ISP for keeping me safe from timing attacks using their unreliable network infrastructure.
At the hardware level any optimizations that change performance characteristics locally (how long the crypto operation directly takes) or non locally (in this case the secrets leak via observation of cache timings in the attacker's untrusted code) are unsafe.
Intel DMPs already have a flag to turn off the same behavior that was exploited on the M1/M2. Which may suggest that the risk of this type of optimization was understood previously.
Mixing crypto operations with general purpose computation and memory accesses is a fragile balance. Where possible try utilizing HSMs, yubikeys, secure enclaves - any specialized hardware that has been hardened to protect key material.
Are there any circumstances where this hardware is accessible in the browser? As I understand, it is not generally available (if at all) for any cryptography you might want to do in the browser.
One thing to remember here, though, is that there isn’t anything special about key material in this attack other than it being a high-value target. If we move all crypto to purpose-made hardware, someone could just start trying to target the messages to/from the crypto system.
This is one of the technical advantages of a blockchain-based system. As long as the keys are protected and signatures are generated in a secure environment, then the content of the message doesn't need to be secret to be secure.
It's not a solution to situations where privacy is desired, but if the reason for secrecy is simply to ensure that transactions are properly authorized (by avoiding the leakage of passwords and session information) then keeping the signature process secure should be sufficient even where general secrecy cannot be maintained.
Why only cryptographic applications? What if I'm writing a very sensitive e-mail, for instance?
In contrast, the timing of virtually any email operation is not dependent on the contents of the email, other than the size. That is, whether you wrote "my password is hunter2" or "my password is passwor", the timing of any operation running on this email will be identical.
What about spell checkers etc? Or even just whatever runs to figure out where to break the lines?
Consider that those programs are not making any effort whatsoever to run in constant time, and yet no one has shown any timing attack against them. OpenSSL has taken great pains to have constant execution time, and yet subtle processor features like this still introduce enough time differences to recover the keys.
That's plausible, but a very different argument from the original, that read:
> In contrast, the timing of virtually any email operation is not dependent on the contents of the email, other than the size.
I’ve been using the TPM 2.0 chip on my ASUS based Linux box to store various keys. Tooling for this on the Linux side has improved significantly [0] and it’s been supported since kernel 3.20 (2015) [1].
How effective this is at improving one’s security posture is another question and it’s probably not a huge security upgrade, but it does mitigate some classes of attack.
I’m curious why you’re saying it’s explicitly not allowed? At least for standard TPM 1.2/2.0 chips, that isn’t the case.
- [0] https://wiki.archlinux.org/title/Trusted_Platform_Module
- [1] https://www.phoronix.com/news/Linux-3.20-TPM-2.0-Security
https://developer.apple.com/documentation/security/certifica...
https://developer.android.com/privacy-and-security/keystore#...
But I agree in general with your point
The questions is what could be the solution going forward, which is going to be a huge change anyway. I do not see a way out of this with our current architectures.
Lets assume I multiply two 64bit numbers. The CPU could just do it the same way every time and the worst-case has 4 cycles latency. It may also track if one of the factors is zero and dynamically replace the multiplication with a zeroing idiom that "executes" in 0 cycles when the scheduler learns that that either input is zero as an extreme example.
Less radical it could track if the upper halves of registers are zero to fast-path smaller multiplications (e.g. 32bit x 32bit -> 64bit) and shave off a cycle. IIRC some PowerPC chips did that, but for the second argument only. The ISA allowed it.
A realistic example are CPUs with data-dependent latency shift/rotate instructions. What do you do if an ISA doesn't specify if shift/rotate is constant time, but every implementation of it so far did it in constant time? Do you slowly emulate it out of paranoia that a future implementation may have variable latency? An other real-world example of this would be FPUs that have higher latency for denormalised numbers its just not relevant to (most) cryptographic algorithms.
How the fuck are you supposed to build anything secure, useful, and fast enough from that?
EDIT: The downvotes make no sense. What this bug has in common with Spectre is that it has to do with cache timing. But in Spectre, the cache is affected by speculative execution; with "GoFetch", it's the pre-fetcher pre-fetching things which look like memory addresses. Pre-fetching is not speculative execution.
But if they did mean to include pre-fetching in "speculation" then I retract my comment