Hacker News Reader

Comment by jhallenworld | Hacker News Reader

I wonder how programmable it is... Assuming it's not programmable would be an argument not to use standard encryption algorithms.

View on HN

No, it would not be. If you're using algorithms or constructions that are physically within the reach of a conventional computer assembled for any amount of money obtainable from the treasuries of all the Five Eyes governments, you're doing something wrong. Don't use RSA-1024; don't use 1024 bit multiplicative group Diffie-Hellman, don't use RC4, and then stop worrying about NSA supercomputers.

Per our previous discussion, you should probably add hash and KDF issues to this list.

I also find supercomputers relevant to PRNGs, both on the overall PRNG security front and on the inadequate seeding front. Since CSPRNGs are not designed to be slow to calculate, supercomputers can be used to attack key generation from inadequate seeding, given some useful model of what the effective space of seeds could have been.

For instance, there are probably still many embedded systems that are initializing without anything that is reasonably called "environmental noise", and their state space is dangerously small, yet maybe still challenging for most organizations to search, depending on exactly what they initialized it with.

The embedded systems cold start entropy problem is a good point I hadn't thought about before.

Of course (not that you disagree), if you have this problem, you have it with or without NSA supercomputers!

I don't think CryptGenRandom or urandom on general-purpose computers is a viable target for this, though, regardless of how "fast" the CSPRNG is.

ASICs are not reprogrammable. And they're expensive to build, but the NSA apparently has the resources to build new ones at semi-regular intervals, so I don't know if switching algorithms would really slow them down that much.

Considering that it takes 1-2 years to spin custom silicon I would say that it is a viable approach.

I think they could probably do significantly faster than that since these would presumably be a relatively low volume, single customer application, and wouldn't need the extensive testing cycle required for mass production ASICs. Which is to say it doesn't take a year from delivering a mask set to a fab to getting packaged parts.

That's not true at all. Were you paying attention to the development of bitcoin asics? Those things popped up fast once the idea spread.

Gate arrays can be pumped out in weeks to days if you control the fab.

I think grandparent was suggesting custom encryption algorithm (potentially less secure) to prevent use of a brute force approach that's custom made for a specific algorithm.

I thought I heard some presentation about superencipherment with AES + some cipher (maybe derived from AES) with per-session randomized S-boxes. You would include the S-boxes in the message as a sort of salt.

The goal of this is that if there is a cryptanalytic attack that gives brute force a 2⁴⁰ speedup and hence attackers build custom hardware to implement it, their hardware is still not effective against the other cipher and they have to do something different (maybe hard to characterize how difficult the different thing would be anticipated to be).

Don't ever do this. The interaction of different crypto algorithms can be counterintuitive (in the sense that you won't gain what you think you're gaining), it's extraordinarily expensive in ways unbalanced against defenders, and it introduces new implementation errors that can weaken the whole system.

If you're paranoid about AES, use Chapoly (or just use Chapoly because it's in many ways more convenient to use than AES). Don't build elaborate cascades (all cascades qualify), and don't randomize S-boxes.

It is hard enough getting these systems right when you're playing exactly by the book. If you're designing a cryptosystem and you're not a professional cryptographer, the list of things you should be worried about getting wrong is very long and very scary.

By "don't build elaborate cascades (all cascades qualify)", do you mean "no cascade construction is preferable to any other", or "every cascade is too elaborate to be worthwhile"? (The second implies the first, but the first doesn't imply the second.)

Every symmetric cipher cascade is bad.

So the cascading cypher options (AES-Serpent-Blowfish) for VeraCrypt volumes are less safe than a simple AES encrypted volume?

Serious question. I know nothing about crypto, just assumed "more is better, but slower".

Anything that uses Blowfish since 2010 is incompetently designed (Blowfish has an 8-byte block size).

AES-Serpent is probably not less safe than AES; it's just not necessarily as much more safe as you'd expect.

A much bigger concern than which precise ciphers you're using is which block cipher mode you're operating under; Truecrypt/Veracrypt uses XTS --- like most disk encryption --- which (among other things) isn't authenticated.

The funniest thing about TC/VC's cascades is that the keys are derived from passwords anyways: a giant clunking complicated block cipher cascade resting on top of a low-entry password secret. It's just a silly design.

What would be your ultimate recommendation on storing sensitive data, if not TC/VC?

It's probably easier to use a standard algorithm in a way that can't be attacked by this machine (e.g. use AES-256 with a properly random key) than it is to create a non-standard algorithm that doesn't have vulnerabilities.

Without some secret algorithmic defect, the existence of which would moot the machine and would throw any encryption algorithm into question, no amount of compute that can be assembled under the physical limits of conventional, non-quantum computing can break a 128 bit key.

AES-128 is the norm. No NSA supercomputer should convince you to use AES-256.

Reply on news.ycombinator.com