Key Takeaways
Anthropic's Claude AI found a hidden weakness in a post-quantum encryption scheme in 60 hours — work that took human experts two years to miss.
Anthropic's Claude AI cut the cost of cracking a post-quantum cipher from 2^64 to 2^38 operations in 60 hours, discovering a vulnerability that human cryptographers had missed through two years of review.
"Claude found a symmetry buried in HAWK's math that no human had thought to use," Anthropic wrote in a paper published July 28, noting the model worked autonomously for 60 hours. The company spent roughly $100,000 in API costs per result, with staff then spending several hundred hours verifying the AES finding alone.
The attack targets HAWK, a digital signature system competing in NIST's post-quantum competition. For its smallest key configuration, the computational cost fell roughly 67 million times. The fix — doubling key sizes — eliminates the compactness advantage that made HAWK attractive for blockchain transactions, where signature size directly affects fees.
Bitcoin's current security relies on ECDSA, a different system untouched by these findings. But the clock on quantum-resistant upgrades is accelerating: Google targets 2029 for its migration, and US agencies face a 2031 deadline under federal mandates. Anthropic's results suggest AI may find flaws faster than humans can verify them, compressing the timeline for cryptographic transitions.
HAWK was one of the finalists in NIST's post-quantum signature competition, selected for its compact keys and fast signing — properties that made it a natural candidate for blockchain applications where every byte of block space carries a fee. Anthropic's disclosure means NIST must now weigh a public attack on a scheme it was still evaluating.
The trade-off is stark. Signature size is block space, and block space is fees. Any chain shopping for a quantum-resistant replacement is partly choosing on bytes per signature. "Unfortunately, doubling HAWK's key size eliminates many of the reasons making the scheme an attractive PQC signature candidate," Anthropic wrote.
Claude also attacked a 7-round version of AES — the cipher protecting most internet traffic — and produced a shortcut named the Möbius Bridge that made the best known attack 200 to 800 times faster. Full AES-128 runs 10 rounds, so nothing in production use is at risk. The model initially refused the task, telling researchers "there's nothing easy to find" in the most-studied block cipher in existence.
The verification bottleneck may prove more consequential than the findings themselves. Anthropic researchers spent several hundred hours learning enough cryptography to confirm Claude's AES result worked. "The cybersecurity community is now grappling with the fact that language models are able to discover so many bugs that the standard human processes struggle to keep up," the company wrote.
Anthropic also broke 13 rounds of LEA, a Korean national encryption standard, in under an hour on a desktop. The deployed version runs 24 rounds, so no fielded systems are affected. The company plans a workshop with academics in the coming weeks and hinted at more unpublished results.
This article is for informational purposes only and does not constitute investment advice.
