Many years ago I attended an academic cryptography conference and took part in a panel there. At the time, I was the lead engineer at a startup making an enterprise blockchain platform and we (I) had made the controversial choice to not use ZKPs in its design. Instead we were using confidential computing with Intel SGX enclaves. Although CC schemes like SGX do use cryptography extensively, it's all "boring" cryptography like key agreement, AES, signatures and so on. Everyone else there was much more interested in exotic cryptography like zkSNARKs, and thought using it in blockchain protocols was obvious.

So on the panel I was the devil's advocate, being challenged over that choice (it was all very polite). One of my points was that I didn't feel comfortable with ZKP systems because they are both very clever and completely non-recoverable. Any mistake at all leads to catastrophic collapse: not only can people mint money at will, but they can't be caught by construction, not even via post-hoc audits because by design there's no way to audit the system. This is quite different to classical Satoshi-style blockchains where the only thing that can cause such problems is a collapse of the core digital signature algorithm, which is very well tested and understood.

After the conference whilst getting drinks, I was chatting to a ZKP researcher and we got a bit drunk. He told he that he was leaving academia to go work for a blockchain company, so of course I asked if he'd be adding ZKPs to their platform. He laughed and said no, he'd never use his own research for anything. He said these systems generate billions of constraints and if even a single one is wrong the entire system fails.

Technologies like SGX are, cryptographically speaking, ugly ducklings. They rely heavily on proprietary technology and security through the obscurity of nanometer scale electronics. But the whole secure hardware world has gone through many rounds of intense combat and learned to build in extensive renewability features. Errors in the implementation are expected and designed for, with many ways to re-seal the system after a compromise and to audit that counterparties have applied the necessary updates. Intel's implementation of all this got a bad rap after Spectre attacks were discovered, but I think that in reality they did well: unlike AMD's implementation which experienced catastrophic collapse several times requiring new hardware, Intel were able to repeatedly patch and reseal SGX in the wild without needing hardware replacements - even against Spectre, which nobody anticipated and goes to the core of CPU design. Additionally, the way I used it in the design of Corda meant that a failure of SGX just led to privacy failures but not logic failures: you couldn't mint money by attacking it, just leak data. But that wasn't the only privacy feature so it was a pure upgrade.

A few years later there was a very similar attack to this one on Zcash. They accidentally published some values from the setup procedure they shouldn't have, and it could be used to forge proofs. Someone told me later that in their view Zcash should have just shut down after that, because the social contract had been violated. Nobody could ever be sure that someone hadn't spotted the problem and minted themselves an unlimited supply of coins.

So we can't say this attack on dusk-plonk is terribly surprising. It's exactly the scenario the researcher warned me about years ago and very close to one that happened a few years later. These algorithms are so insanely complicated that even the researchers that write the papers - in academia, under no time pressure at all - can't implement them correctly! If the original researchers can't do it then what chance do other people have? I worked on Bitcoin for years but I'd have serious reservations about using a coin that relied on ZKPs because I could never have confidence that the money supply was secure.