An AI model just cracked a cryptographic algorithm that survived two years of expert human review in just 60 hours, and the implications for Bitcoin and the entire crypto ecosystem are profound.
By Keisha Williams | July 30, 2026
The Hook
Anthropic, the AI company behind the Claude assistant, announced this week that its Claude Mythos Preview model discovered a serious flaw in HAWK, a proposed post-quantum digital signature scheme that was being considered as a replacement for the cryptographic tools that currently protect online banking, web payments, and — eventually — cryptocurrency transactions.
The AI-driven attack took roughly 60 hours and about 100,000 USD in computing costs. It effectively halved the smallest key strength of HAWK, reducing the work needed to break it from approximately 2^64 operations down to 2^38 operations. For context, that is the difference between “practically impossible to crack” and “theoretically within reach for a well-funded attacker.” HAWK had survived two years and two rounds of expert human review before the AI found what humans could not.
On-Chain Evidence
Digital signatures are the invisible foundation of cryptocurrency. Every time you send Bitcoin, the network verifies your right to spend those coins using a digital signature scheme. Every time Ethereum processes a smart contract transaction, signatures prove who authorized it. These cryptographic tools also protect every website showing a padlock icon in your browser — they are the plumbing beneath the entire digital economy.
The signature schemes used today by Bitcoin and Ethereum are expected to eventually fall to quantum computers. That is why organizations like the National Institute of Standards and Technology (NIST) have been running multi-year competitions to select new “post-quantum” algorithms — including HAWK — that can withstand attacks from both classical and quantum computers.
Anthropic’s AI did not break Bitcoin or Ethereum directly. Both networks currently use elliptic curve signatures, which neither the HAWK attack nor the AES improvements directly targeted. HAWK has not been deployed anywhere in production. But the speed and efficiency of the AI-driven attack sent a clear warning: the cryptographic assumptions underpinning digital assets are eroding faster than anyone expected.
The Core Conflict
The timing of Anthropic’s discovery is jarring for the crypto world. BIP-360, the proposal to give Bitcoin quantum-resistant addresses, specifies three algorithms NIST has already standardized, with multiple options included deliberately so users have fallbacks if one is later broken. BIP-361, the companion proposal that would freeze vulnerable coins — potentially more than a third of Bitcoin’s supply — argues that the migration window is closing because cryptographic attacks are improving by up to 20-fold.
Anthropic’s results align with that trend perfectly. What took human cryptanalysts years to evaluate, an AI model accomplished in days. And it is not just HAWK. The same Claude model produced improvements against a deliberately weakened version of AES — the cipher used to encrypt wallet files across the industry — by factors of 200 to 800. It also found smaller improvements, under tenfold, against Poseidon, the hash function that secures many zero-knowledge proof systems used in rollups and privacy protocols.
The privacy network Zcash activated its Ironwood upgrade just this week, introducing a new shielded pool designed to remain secure even if quantum computers arrive. The upgrade retires the old Orchard pool, where a counterfeiting bug sat undetected for four years. Every coin in Zcash’s roughly 80 million USD new Ironwood pool must now exit through gates that cap withdrawals at verified deposits — a direct response to the kind of existential cryptographic risks Anthropic’s research highlights.
Market Implications
For everyday crypto investors, the practical concern is not that an AI will break Bitcoin tomorrow. It will not. The real question is whether the Bitcoin and Ethereum communities can upgrade their cryptographic foundations fast enough to stay ahead of both quantum computers and AI-assisted classical attacks. The BIP-360 and BIP-361 proposals show that Bitcoin developers are taking the threat seriously, but consensus in a decentralized network is slow by design.
Investors holding coins in addresses generated years ago — including the estimated 1.1 million Bitcoin believed to be held by Satoshi Nakamoto — face the greatest exposure. Those early addresses use signature schemes that would be first to fall if a sufficiently powerful quantum computer comes online before the network upgrades. BIP-361 would allow the community to freeze those vulnerable coins, but that raises profound questions about who gets to decide which coins are “abandoned” and which are simply held long-term.
For projects building on zero-knowledge proofs — including Ethereum rollups, privacy coins, and identity systems — the Poseidon improvements are a reminder that even cutting-edge cryptography is not immune. Developers relying on ZK proofs for billions of dollars in daily transactions need to monitor these developments closely and be prepared to rotate to alternative hash functions if the attacks improve further.
The Verdict
Anthropic’s discovery is a wake-up call wrapped in good news. The attack on HAWK was found through responsible disclosure — Anthropic notified the algorithm’s authors in June and coordinated publication with NIST’s public mailing list. No deployed system was harmed. But the pace at which AI is now able to find cryptographic weaknesses suggests the entire timeline for post-quantum migration needs to accelerate.
For investors, this is a long-term story with short-term implications. The networks that move fastest to upgrade their cryptography — and that have governance structures capable of making hard decisions about vulnerable coins — will be better positioned to survive the coming decade. Those that drag their feet are betting that AI and quantum computing progress will be slower than every sign suggests it will be. That is not a bet anyone should want to lose.
Interestingly, Claude initially refused the AES problem, telling researchers it was “genuinely hard” and that there was “nothing easy to find.” After three short prompts, it produced a billion output tokens over three days and found the improvement. Verifying it took two Anthropic researchers nearly a month. The lesson: AI might not always get it right on the first try, but when it does find something, humans had better pay attention.
The cryptocurrency market remains highly volatile. This article is for informational purposes only and does not constitute financial advice.
Disclaimer: This article is for informational purposes only and does not constitute financial advice.
60 hours and 100k to break something humans couldnt crack in 2 years. post-quantum crypto just got a serious wake up call
reducing 2^64 to 2^38 is not a small tweak, thats a catastrophic break. NIST needs to redo the HAWK evaluation ASAP
if Claude can do this in 60 hours what happens when someone points it at wallet software? the attack surface is terrifying
2^64 down to 2^38 in 60 hours for 100k. nist is probably rethinking their whole vetting process right now
26 bits of security gone for 100k in compute. every finalist in the nist queue gets re-run through adversarial models now, no way they wait for the old schedule
two years of human review missed this and claude found it over a weekend. post quantum crypto is going to need post-human review apparently
anthropic really out here doing the NSA job better than the NSA lol
Funny until you realize a single company now holds that capability. Nobody elected a lab to be the industry cryptanalysis referee, and the next discovery might not come with a blog post.
This is exactly why Bitcoin conservatism with SHA-256 matters. You do not want to be rushing into new schemes that have not been battle-tested against AI fuzzing.
100k in compute to break a scheme that survived two years of expert review. the math on that ROI is brutal if you think about what a hostile state actor with actual budget could do
the 2^38 number is the scary part. thats well within reach of a botnet or a funded attacker. HAWK is done, nist needs to pull it
segfault_eth the 2^38 figure is scary but HAWK was still in the NIST review pipeline. this is exactly why the process exists. better to break it now than after deployment
people panicking about AI breaking BTC. read the article. HAWK isnt deployed anywhere and current BTC sigs werent touched. breathe
You’re right that HAWK isn’t deployed yet but that’s not the point. The point is AI found in 60 hours what 2 years of human cryptanalysis missed. What happens when someone runs this against deployed schemes?
Daniel K. the real question is whether Claude can find similar flaws in deployed schemes. HAWK was new and untested but what happens when someone runs this against Ed25519
sha256 and secp256k1 are a different class of hard though. lattices were always the fragile corner of post quantum. still, agreed the wake up call is real
HAWK was in NIST review, not deployed. but the speed of the AI attack is what changes everything long term
Correct on BTC, but the comfort only lasts until someone points the same pipeline at the PQ migration everyone is supposed to start. The reassurance has a shelf life.
60 hours of compute vs 2 years of human review. every zk team should be running AI fuzzing on their circuits yesterday
Petra K. running AI fuzzing on zk circuits is table stakes now. any team not doing it after this paper is negligent
agreed but who pays for the fuzzing. 100k in compute per candidate adds up when NIST has a whole pipeline of schemes to grind through
Two years of expert review dismantled in 60 hours. Every audit budget in the industry just became a rounding error against a hostile model run.
the scary version isnt claude finding HAWK. its whoever runs the same sweep and doesnt tell anyone. bug disclosure only works when the finder wants it to
2^64 to 2^38 in 60 hours for 100K in compute. the economics of breaking crypto just changed permanently
permanently is doing heavy lifting there. one broken NIST candidate doesnt mean every lattice scheme falls to a 60 hour run. HAWK had a specific structural weakness