A debate from 2013, a $95 search bill, and what a 38-round collision really means
SHA-256 · SHA-1 collision · cryptography · Bitcoin · November 2026
The argument we never settled
In 2013, Arindam, Josh, Rami and I spent a while debating a question that sounds like trivia and is really about trust: how long did SHA-1 have left? As I remember it, nobody doubted that someone would eventually break it. By then it had already been shown breakable on paper. The argument was about when, and what it would cost.
This week I sat down to turn that memory into a LinkedIn post, and I did what I should always do before posting a fun fact: I checked it. My version went like this: in 2017 Google hacked SHA-1 for $100,000, and in 2015 somebody cracked SHA-256 at the “38th attempt”. Parts of that are right. Parts are wrong in ways that turned out to be more interesting than the story I remembered.
First, what are we even attacking?
Picture a machine that takes any file, a photo, a contract, a software update, and stamps out a short fingerprint. Change one letter in the file and the fingerprint comes out completely different. Nobody can run the machine backwards to recover the file. SHA-1 and SHA-256 hash functions are two such machines. They sit quietly behind software updates, digital signatures and website certificates, and SHA-256 also underpins Bitcoin.
The whole point is that two different files should never share a fingerprint. When they do, that is called a collision, and it is the crack a forger needs: show you a harmless file to approve, then swap in a different one wearing the same stamp. Security people do not wait for this to become easy. They watch for the first time it is merely possible.

Figure 1. The two fingerprint formulas have followed very different paths.
SHA-1: a twelve-year countdown
In 2005, researchers showed that SHA-1 could, in theory, be beaten with far less work than blind guessing. “In theory” covered a lot of ground. The attack would still need a computing bill that no ordinary criminal would pay. So the industry did what industries do with a warning that is not yet urgent: it kept using SHA-1 and promised to migrate later.
By 2015, the researchers at CWI in Amsterdam were warning that a real collision could be bought for roughly $75,000 to $120,000 of rented cloud graphics cards. That is the number that stuck to the story, and it is where my “$100,000” came from.
Then, on 23 February 2017, Google and CWI stopped warning and showed it. They published two PDFs that look nothing alike but have the same SHA-1 fingerprint. CWI’s own example of the danger was two contracts with different fees and one fingerprint. The effort was about nine quintillion SHA-1 calculations. Google described it as 6,500 years of one processor plus 110 years of one graphics card, spread across a huge fleet of machines. If every person on Earth had done one calculation per second, it would have taken about 36 years.
Two details are worth keeping. The attackers built both files together, so they could not take a document that already exists and forge a twin of it. And neither Google nor CWI sent anyone a bill, so “a $100,000 hack” is shorthand, not an invoice. CWI’s headline put the real lesson in six words: “Industry deprecation proved to be too slow.”
SHA-256: the 64-round lock
SHA-1’s stronger successor is SHA-256, and it is what most of the modern internet leans on. Inside, it scrambles data in 64 rounds, each one mixing up the output of the last. Cryptographers have a habit for testing a lock like this: they do not try the whole thing first. They attack versions with some rounds removed, like picking a lock with several pins taken out, and see how many pins they can handle. The results work as an early warning system.
In 2013, the same year as our debate, researchers Mendel, Nad and Schläffer published the state of the art: a practical collision on 28 rounds, a collision on 31 rounds, and a flashy result on 38 rounds with a big asterisk. The asterisk was that they got to choose the starting state of the function, something no real attacker can do. Cryptographers call it a semi-free-start collision. The plain version is a head start, like a sprinter who gets to pick where the starting line goes.
That asterisk is where my “38” came from. I suspect it is also where my “2015” came from, because a copy of the paper sits in a public archive under the number 2015/350.
The plot twist
Then, this autumn, a researcher named Omer Goldzweig started posting on X. On 8 October he put up two messages and said they collide for 37-round SHA-256, with a paper to follow. On 2 November he posted two more and wrote that this one “was a little harder”: 38 rounds.

Figure 2. The announcement on X, 2 Nov 2026 (screenshot of the public post by @omergoldz).
The paper arrived on 15 November. The round count is only half of what makes it interesting. The other half is the price tag. By the author’s account, the search ran on rented graphics cards for about 53 hours, around $95, and got lucky: it finished after roughly a tenth of the work the math predicts. The expected bill is closer to 440 GPU-hours, which by my arithmetic from the paper’s own numbers is still under $1,000. Before the hunt there was a one-time preparation of about 52,300 processor-hours (around $290) and a 7-gigabyte lookup table.
Why does that matter? Because the earlier published estimates for the same job were astronomically larger. According to the paper, the best previous estimate for 38 rounds was around 2^104 operations, a number with 32 digits. This attack needs about 31 quadrillion, a 17-digit number. That is a gap of roughly 800 trillion times, though to be fair, the earlier figures were estimates on paper rather than attacks someone had actually run.
How? In rough terms, the author fixes most of the second half of the message in advance and builds lookup tables, so the conditions in the early rounds are met by a table lookup instead of by luck. The genuine gamble is saved for the later rounds. It is a locksmith who prepares a ring of nearly right keys, so each real attempt only has to get lucky on the last few pins.
I didn’t take his word for it
Here I owe you a disclosure. I used Claude, an AI assistant, to help check this and write it up. And the paper itself states that it was predominantly written by AI agents, mainly Claude Opus 5.5, under the author’s direction. An AI-assisted claim of a cryptographic milestone deserves a cold, independent check, so I had Claude build SHA-256 from scratch with an adjustable number of rounds. It first confirmed that the code gives the right fingerprint for the test input “abc” at the full 64 rounds. Then I fed it the two messages from the tweet. They match the paper’s table exactly.

Figure 3. The two messages are identical in 12 of 16 chunks, yet end up with the same fingerprint.
The result held up. At 38 rounds, the two different messages produce the same fingerprint. At 39 rounds, and at the full 64, they do not.
Watching the rounds go by is oddly beautiful. The two runs start out identical. They split apart at round 8 and stay different through round 24. Then, as if choreographed, the differences cancel and the two runs fall back into lockstep at round 25 and stay together through round 38. At round 39 the dance falls apart. The attack is engineered so that the differences cancel at just the right moment, and the paper says that trick does not stretch to 39.

Figure 4. The match exists at 38 rounds and disappears at 39.
So is SHA-256 on the clock?
No, and the people to listen to most closely are the ones who are excited. There are three reasons to stay calm.
But there is a reason this one stuck in my head. SHA-1’s countdown ran twelve years, from “broken on paper” in 2005 to “broken in practice” in 2017, with the warning light visible to everyone the whole time. What just changed for SHA-256 is the price of the 38-round milestone: from “out of reach” to “under a thousand dollars of rented cloud time, on average”. That tells us how fast the frontier is moving. It is not a forecast about round 64.
For ordinary people and companies, there is nothing to change today. SHA-256 remains the standard, and SHA-1 is the one that should already be retired.
Back in 2013, the question was when. For SHA-1, the honest answer turned out to be “sooner than the industry wanted”. For SHA-256 the honest answer today is “not yet”. The people watching the clock are right to keep watching it.
Related on Lots of Data: AI-Enabled Cyberattacks · Find, Fix, Repeat (Patch Calendar) · Riemann Hypothesis & Encryption
