AIBRIEF
Johns Hopkins cryptographer Matthew Green says AI assisted codebreaking is arriving during the same decade NIST is replacing the public key math behind logins, bank transfers, and encrypted messages.
A working cryptographer says the global swap of internet cryptography is running on the same clock as AI's improving ability to challenge it, and that collision, not any single result, is the story worth watching.
On 29 July 2026, Johns Hopkins cryptographer Matthew Green published a blog post arguing that the moment an AI gets good at cryptanalysis research is the worst possible moment for the public-key math that secures logins, bank transfers, and encrypted messages to still be in mid-replacement. He tied that worry to Anthropic's research release describing AI-assisted key recovery against HAWK, a post-quantum signature candidate now under NIST review, plus a separate "Möbius Bridge" analysis of AES.
Nothing in the release shows a deployed cipher broken, and Green is careful to say so. HAWK is one of many candidates, not a finalized standard, and the reproducible demo sits alongside a self-reported figure of roughly $100,000 in compute. An independent benchmark paper frames how to evaluate LLM-driven cryptanalysis. Green explicitly references Impagliazzo's "Minicrypt" as a worst-case world and offers a best-case read where the literature becomes more robust.
What remains unknown: whether the HAWK attack holds up under independent review, and whether AI cryptanalysis improves fast enough to land before NIST finalizes the next batch of standards.