An AI Model Just Killed a Post-Quantum Cipher — and Factored Two RSA Records in 16 Days
On July 28, 2026, a post to NIST's public cryptography mailing list ended a signature scheme. The next morning, its designers withdrew it from the standardization race. The attack that killed it was found by an AI model.
On July 28, 2026, a post to NIST's public cryptography mailing list ended a signature scheme. The next morning, its designers withdrew it from the standardization race. The attack that killed it was found by an AI model.
That sentence still sounds strange when I read it back. I build an encrypted messenger, so I spend a lot of time thinking about which algorithms to trust and for how long. This story changed how I think about both questions, and not in the way the headlines suggest.
What actually happened to HAWK
HAWK was a candidate in NIST's second post-quantum signature competition, the one launched in 2022 to find more signature options beyond the standards finalized in 2024. It was a lattice-based scheme, built on a problem called the lattice isomorphism problem, and it had survived two rounds and about two years of expert review. In May 2026 it advanced to round three, one of nine candidates left standing. It was the only lattice scheme among them.
Its selling point was compactness. HAWK signatures were small and fast, and unlike Falcon, the other compact lattice signature people talk about, it avoided floating-point arithmetic entirely. That matters more than it sounds like it should. Floating-point code is where subtle, platform-dependent bugs live, and in cryptography those bugs become vulnerabilities.
Anthropic's Claude Mythos Preview found a symmetry in HAWK's lattice that nobody had noticed. The model worked for about 60 hours, at roughly $100,000 in API cost, inside a harness running several agents at once. One agent said the attack was impossible. Another one found it anyway. The human steering the work had a background in theoretical computer science but no expertise in this particular kind of cryptography. Anthropic describes his contribution as project management.
The attack recovers a nontrivial automorphism from a public key and uses it to build an equivalent signing key. For HAWK-512, the smaller of the two proposed security levels, the estimated cost of stealing a key fell from about 2^150 operations to about 2^108. Each step down halves the work, so that drop is a factor in the trillions. Anthropic published working code that recovers a key for HAWK-256, the deliberately weakened test variant, in about three hours and 42 minutes on a 96-core server.
Here is the part that tells you how cryptography actually works as a discipline. Anthropic shared the attack privately with HAWK's designers in June. On the afternoon of July 28, Stephen Weis posted it to the pqc-forum mailing list with the code. That same evening, cryptographer Daniel Apon replied that he had checked the reduction independently and it held up. The next morning, Léo Ducas, a lead designer of HAWK, wrote on the team's behalf that the obvious fixes, doubling the parameters or moving to higher-rank modules, would make HAWK uncompetitive, and that the team would withdraw it. He thanked Anthropic and wished the remaining candidates well.
Two days, in public, on a mailing list anyone can read. That is the system working. It is also the system working much faster than it used to.
The AES result, and why it is not what it sounds like
Mythos produced a second result the same day: an improved attack on a seven-round version of AES-128. The model invented a technique it called the Möbius Bridge, a meet-in-the-middle trick that removes a 256-way guessing step, and the attack runs 200 to 800 times faster than the previous best.
Read that again carefully, because the number is designed to mislead you. Seven rounds, not ten. AES-128 uses ten rounds, and researchers routinely study weakened versions as practice targets to measure how much safety margin remains. Even against the seven-round version, the attack still needs about 2^105 chosen plaintexts. That is not a number you can collect. It is a number you can write on a whiteboard.
So no, AES is not broken, your chats are not readable, and nothing needs patching. What changed is that an AI model did original cryptanalysis work on one of the most studied ciphers in history and produced a technique nobody had published before. That is a different kind of news, and I think it is the more important one.
Then September happened
Six weeks later, the classical factoring record, which had not moved since February 2020, broke twice in sixteen days. Both times the work was done by AI agents porting old code to GPUs.
On September 3, Eric Lu at Cognition announced the factorization of RSA-260, a 260-digit number, 862 bits. His team used Devin agents to rewrite CADO-NFS, the open-source implementation of the general number field sieve, for GPUs. The run took about 4,900 GPU-days on idle nodes of an LLM training cluster and cost roughly $400,000 at market prices. Lu says that is about ten times cheaper than the previous public level. He also says, plainly, that he does not understand the mathematics in depth. His job was setting priorities, running benchmarks, and catching mistakes in time.
On September 19, Stephen Weis at Anthropic factored RSA-896, an 896-bit number, with Claude. Up to 2,048 H100 GPUs, about ten days, roughly 30 GPU-years of compute. Weis says he did not meaningfully improve the running time of the number field sieve. He also says RSA-1024 keys are now within reach of anyone with a data-center-scale GPU fleet.
Put the two together and Cognition's estimate lands at about $30 million to factor a 1024-bit RSA key in rented GPU time, over five or six months. That is an estimate, not a demonstrated fact. RSA-1024 has never been publicly factored. But $30 million is nation-state money, not science-fiction money, and it is a number that was $40 to $120 million and one to two years before September.
Two details keep me honest here. First, there was no algorithmic breakthrough in either record. The general number field sieve is from the 1990s. The advance was engineering: getting a thirty-year-old algorithm to run well on GPUs, and doing it with agents that could iterate for weeks while a human checked in every couple of hours. Second, RSA-2048 is roughly a billion times harder than RSA-1024. Your TLS connections and your 2048-bit keys are not in danger from this.
What this means if you actually build things
I keep coming back to three lessons, and none of them is "panic."
The first is that the review clock got faster. HAWK survived two years of expert review and died in two days once a model looked at it with fresh eyes and enough compute. That is good news. It means the standardization process works the way it is supposed to: candidates get attacked in public before anyone deploys them. But it also means a scheme's track record is worth less than it used to be. Two quiet years no longer means what they meant. If you are picking algorithms for something you will ship, the question is not only "has anyone broken this" but "has anyone with a serious model and a serious budget tried."
The second is crypto agility, and I mean it in the boring, practical sense. The finalized standards are untouched. ML-KEM, ML-DSA and SLH-DSA rest on different assumptions, and Anthropic says the attack does not extend to them. That is why my own messenger uses ML-KEM-768 for key exchange and ML-DSA-65 for identity signatures: standardized, reviewed, boring. But HAWK was the only lattice option left in the additional-signatures on-ramp, and now it is gone. If you built on a candidate instead of a standard, you just watched your foundation get withdrawn in a morning. Design your code so the algorithm is a plug, not a load-bearing wall. The next withdrawal will not give you more notice than this one did.
The third is the unglamorous one: retire RSA-1024 now. Not because anyone will factor your key tomorrow, but because there is no reason left to keep using it. Both apple.com and microsoft.com still publish 1024-bit RSA keys for DKIM email signing, which anyone can check in DNS right now. A DKIM key signs email, and email lives forever in archives, so a key that becomes factorable in five years retroactively forges every message it ever signed. Rotating to 2048-bit keys costs an afternoon. Keep the 1024-bit key and you are betting that $30 million stays out of reach forever. That is a bad bet to make on someone else's behalf.
There is a fourth lesson, and it is the one I am still turning over in my head. Anthropic's own specialists needed nearly a month to understand the method Mythos proposed for HAWK. The model found the attack, and then humans had to work for weeks to verify that the attack was real and correct. Verification, not generation, is becoming the bottleneck. That flips the usual story about AI in security. The scary part is not that models write exploits faster. It is that they now produce cryptanalysis that takes experts a month to check, and the checking is what keeps the whole system honest.
The people who break ciphers and the people who verify the breaks are now running at different speeds. That gap is where the next few years of cryptography will be decided, and I would rather be building on the side that assumes the gap exists.
Sources: Anthropic, "Discovering cryptographic weaknesses," July 28, 2026; NIST, "Round 3 Additional Signatures," updated July 29, 2026; Cognition, "Factoring RSA-260," September 2026; Stephen Weis, "RSA-896 factored," September 19, 2026; The Wall Street Journal, "Quantum Computers Are Poised to Crack Encryption. AI Could Get There First," November 6, 2026.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.