SHA-256 Collision Hits 38 Rounds: Is Bitcoin Safe?

Bitcoin coin beside hash digits illustrating a SHA-256 collision

Cryptographers have a new SHA-256 collision, and it arrives with a price tag of roughly US$95. It only covers the first 38 of the 64 rounds that Bitcoin relies on, so the network is not in danger today. Still, the speed of recent progress makes it worth a look.

Omer Goldzweig’s preprint reached the IACR archive on 16 November. It gives two 128-byte messages that hash to the same value when SHA-256 is stopped after round 38. Add a 39th round and the outputs share nothing. Earlier work at this round count produced estimates only, starting with 2^104.3 operations in June. The new paper puts the cost at 2^54.8 and says the search needed 53 GPU hours, helped by an early hit at one tenth of the expected effort. It also states that Claude agents did most of the work under the author’s direction.

How the 38-round trick works

Re-running the pair shows the match at rounds 1 to 7, where the messages are still identical, and again at round 38. Nowhere else.

The two messages disagree in words 7, 8, 10 and 15 of their sixteen. Later words are computed from earlier ones, so the attacker has no say over them. The differences are built to cancel in the hash’s internal state by round 24, and from there every computed word has to come out the same in both runs. That holds through word 37 and breaks at word 38.

So fourteen rounds of the collision cost nothing extra. The word schedule just stays quiet for that long with these differences. Stretching the idea to 64 rounds would need 39 quiet words in a row, and the paper itself says more rounds would need a different approach.

Why miners and holders can relax

Mining needs a header whose double hash lands under a target. A collision gives a miner nothing there, since that task depends on preimage resistance. Attacks on that property are far weaker: 45 rounds held as the best from 2012 until a paper at this year’s EUROCRYPT reached 47, and the 45-round attack saves only about 30% over guessing.

A complete collision would not touch confirmed transactions, mined blocks or funded addresses either. An attacker would control both colliding inputs and couldn’t match a hash somebody else already published.

The real exposure is in shared setups. A P2WSH address is the hash of a script. Someone who helps write that script could craft a second one with the same hash that pays only them, wait for others to fund the address, then spend through the second. Signature hashes allow the same trick, and one identifier could stand for two different blocks or transactions shown to different parts of the network.

Bitcoin’s developers saw this coming. Older addresses use a 160-bit hash with only 80 bits of collision strength, and BIP 141 moved scripts to 256 bits. Reaching 2^128 would take about ten billion years at the current 1.04 ZH/s network rate, which works out to 2^94.7 header hashes a year, and mining chips can’t run a collision search at all.

A bounty nobody has claimed

Since September 2013, addresses set up by Peter Todd have paid anyone who submits two different messages with the same hash. One address exists per hash function and no signature is required.

The SHA-1 version was drained on 23 February 2017 at 13:08 UTC, the day Google and CWI released the first SHA-1 collision. That claim took 2.48 BTC. The SHA-256 address holds 0.27734251 BTC across ten unspent outputs as of block 977,036, including 0.12572363 BTC added later that same day. Messages are capped at 520 bytes. This paper’s are 128.

Does the trend point to 64 rounds?

On paper, yes. Real collisions went from 28 rounds in 2013 to 31 in December 2024, then 35, 36 and 38 this year. The estimated cost at 38 rounds dropped by about fifty bits between June and November.

That reading is shaky. A 2013 paper with the 28-round record also described a 38-round collision for an attacker allowed to pick the starting state, something the real function forbids, using differences in the same four words. The new result mostly turns a thirteen-year-old theoretical frontier into a working one, and that frontier has crept just one round further, to 39 in 2024.

SHA-1 offers the grim comparison. A full attack was announced in 2005 at 2^69 and nobody found an actual collision until 2017. For SHA-256, nothing exists below 2^128. The 2^54.8 figure is the paper’s own claim, and as of 22 November there was no public review and no discussion on the Bitcoin developer list.

What to watch

The odd part is the method. An author with no listed affiliation steered AI agents to a result the field’s own estimates placed fifty bits away five months earlier. If that pace repeats, forecasts for 39 or 40 rounds look less reliable as a floor. Expect cryptographers to check the 2^54.8 claim first, then try to extend the same SHA-256 collision approach by a round or two.

Source: 808bits

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.