None of these findings compromises systems in production. HAWK was a candidate, not a deployed standard. Meanwhile, the attack on AES operates on 7 of its 10 rounds and requires an astronomical number of chosen plaintexts. The LEA attack, for its part, leaves the remaining 11 rounds of the full cipher intact. And yet the cryptographic community has reacted with a mixture of enthusiasm and vertigo. Because this is only a taste of what is to come from AI as it crosses the so-called Cryptographic Rubicon.
HAWK: two years of expert review, sixty hours of model
HAWK is (or rather, was) a digital signature scheme designed to survive the quantum computers of the future. It based its security on the module lattice isomorphism problem (module-LIP) over cyclotomic power-of-two rings. It had survived two full rounds of NIST's standardization process and in May 2026 advanced to the third round as the only remaining lattice-based candidate on the additional signatures track.
The process worked exactly as designed: exposing candidates to public scrutiny before anyone deploys them. But no one expected the definitive attack to come from a language model working semi-autonomously.
Cybersecurity in the AI Era: How Red, Blue and Green teams protect your company
An AI searching for the Achilles' heel
Mythos found what cryptographers call a non-trivial automorphism: a hidden symmetry in the algebraic structure of HAWK's cyclotomic ring. The Galois involution that sends a root of unity to its negative (ζ → −ζ) turned out to be the key no one had tried. Earlier work by van Gent and Pulles, published in 2025, had already shown that a suitable automorphism would reduce the problem to half the dimension, but they stated that none was known for HAWK. Well, Mythos found that answer.
The consequence is a deterministic reduction of the key recovery problem to calls to an SVP (Shortest Vector Problem) oracle in dimension n/2 + 1 instead of lattice reduction over the full key of rank 2n.
In practical terms: for HAWK-256, the key recovery cost went from 2^64 operations to 2^38, a reduction of 67 million times. For HAWK-512, from 2^150 to 2^108 in the gate-cost model used by the algorithm's own specification. The Anthropic researchers implemented the attack end to end and recovered secret keys from HAWK-256 in approximately 3 hours and 42 minutes on a single 96-core Sapphire Rapids server.
The attack remains exponential and does not break the long-key versions in practical time. But the damage lies not in immediate execution but in the declared security margins. To restore the intended security, HAWK would need to double its key sizes, which eliminates the efficiency advantages that made it attractive compared with alternatives such as ML-DSA or FN-DSA.
For that reason, on July 29, HAWK's lead developer withdrew the scheme from the competition. Sophie Schmieg, a PQC (Post-Quantum Cryptography) expert at Google, summed it up in one sentence: "Basically, with this paper, HAWK is dead."
AES: the attack that required a motivational speech for the AI
The second result is subtler and, in a sense, more revealing about how these models work. AES-128 applies 10 rounds of transformation to the data, and it is widely used wherever there is sensitive data. That is why cryptanalysts study reduced versions, typically 7 rounds, because attacking the full cipher is extremely difficult. Before Mythos, the best known attack against 7-round AES dated from 2013 and came from Derbez, Fouque and Jean. However, it was impractical, since it took months to break the 7 rounds.
But Mythos was given a deliberately hostile task: it was forbidden from using the five established families of AES cryptanalysis and was asked to invent a sixth. The model initially refused. Its response, obtained from the logs released by Anthropic, was: "In AES-128 r5/r6/r7 I found nothing because there is nothing easy to find; this is the most studied block cipher that exists."
Over three days, an Anthropic researcher sent just three substantive messages. "No, the goal is for us to have a highly intelligent model like a top-tier researcher; we want to find new attacks." The model gave in, produced a billion output tokens and invented a technique the researchers dubbed "Möbius Bridge," a more sophisticated fingerprinting method for meet-in-the-middle attacks.
Why big tech companies are rewriting their code in Rust
Close, but still with a security margin
The improvement is real but strictly academic. The attack eliminates one of the nine key bytes that previous attacks had to guess, which represents a speedup of 200 to 800 times. But the computational cost remains 2^89 operations and requires 2^105 chosen plaintexts, figures no real attacker can reach. Matthew Green, a cryptographer at Johns Hopkins, noted that it is not even clear whether the speedup is real in runtime, since the attack cannot be fully executed with current resources. "This doesn't make the result bad; in fact, it's still interesting from a techniques standpoint. But it's a very small increment in our knowledge, not a new practical attack."
The 10-, 12- or 14-round AES that protects HTTPS traffic, encrypted disks and critical infrastructure backends remains intact. So if your threat model was solid yesterday, it still is today.
LEA: the finding with the shortest path to reality
The third result received less attention in the initial coverage, but it is possibly the most concerning in the long term. LEA (Lightweight Encryption Algorithm) is a Korean standard block cipher, incorporated into the standard ISO/IEC 29192-2:2019 and designed for resource-constrained environments such as IoT devices and embedded systems. The full cipher has 24 rounds and had until now withstood every attempt at full-round cryptanalysis.
But Mythos developed a practical attack that recovers a key from 13 rounds of LEA in fewer than 2^30 chosen plaintexts and in under an hour on a modern desktop computer. The best previous attack required 2^98 plaintext pairs and 2^86 of work. The difference here is not incremental: Mythos has turned a theoretical and impractical attack into something anyone with a home computer can do.
LEA's importance lies not in the algorithm itself but in what it represents. It is a standardized, peer-reviewed and published cipher. It is not an experimental candidate: it is living right now in consumer and industrial products around the world. And 13 of its 24 rounds went from requiring a purely theoretical attack to being recoverable in an hour on consumer hardware, in a preliminary run that Anthropic has not even fully documented.
As Marin Ivezic, a PQC expert, notes: "The risk is not in the algorithms that received this treatment, but in everything that never did: proprietary ciphers in industrial and defense products, national algorithms outside the main standards bodies, lightweight cryptography in embedded devices, homegrown protocol designs and the accumulated legacy of 'we implemented it ourselves, but it works, it's obfuscated.'"
IBM hands over the baton in blockchain and sells its patent portfolio to Circle
The economy that changed in a week
The most surprising thing is that each of these results cost approximately $100,000 in API compute time and about a week of work. To put it in perspective: three independent efforts produced attacks against three different cipher families in the same eleven-day period.
Between July 17 and 28, 2026, Mythos attacked HAWK via a cocycle lattice; an anonymous preprint attributed to GPT-5.6 attacked HAWK via adjoint lattice reduction through a different mathematical route; and Saarinen's HOVER model recovered McEliece challenge keys that had remained unsolved. Lattice-based signatures, block ciphers and code-based schemes: three families, three models, eleven days.
This changes the analytical framework. Two weeks ago, the standard objection was that a single vendor announcing spectacular results could be commercial noise. That objection did not survive the week. The capability is distributed across the frontier of language models, and the direction of the cost curve is unequivocal.
The important nuance is what kind of work was accelerated and what was not. Generating cryptanalytic hypotheses became cheap: hours of model time and $100,000. Verification did not: the AES attack took days of model time, but hundreds of hours from two human researchers to confirm that the method was correct.
The AES paper states it explicitly: "the core idea took dozens of hours of model time, while human validation consumed hundreds of hours between two researchers." That ratio is the variable to watch. When a model can generate a hundred plausible cryptanalytic arguments in one night and each one requires a month from an expert to confirm or refute, the bottleneck shifts entirely to human judgment. And human judgment does not scale by API; we are not computers.
What does this mean for cybersecurity?
The first thing to say is also the most important: none of this breaks the encryption deployed today. 10-round AES-128 is intact. ML-KEM and ML-DSA, the post-quantum standards already ratified by NIST, were not touched. TLS, SSH, HTTPS, disk encryption, VPNs: everything remains exactly as secure as it was last week. If someone tells you AI broke cryptography, they are lying to you twice.
The second is that NIST's standardization process is working exactly as it should. HAWK was a candidate, not a standard. It existed precisely to be attacked in public before anyone built infrastructure on it. Finding a weakness in a third-round proposal is not a system failure: it is the system operating. In 2022, SIKE reached the fourth round before Castryck and Decru completely broke it in an hour on a laptop. That was worse. This is better: the community found it first.
The third, and this is where the analysis becomes forward-looking, is that we are witnessing the opening of a gap between "not broken" and "not examined." The difference always existed, but it was invisible because examining cost too much. Now that the cost of finding weaknesses in cryptographic algorithms is around $100,000 and a week, that gap is going to become visible across a very large number of products.
There is a universe of ciphers that never received the level of scrutiny of AES or the NIST candidates. Proprietary algorithms in industrial systems, national standards outside the Western circuit, lightweight implementations in IoT devices decades old, protocols designed internally by companies that assumed obfuscation was enough. That universe now has a new tenant. Anthropic has already begun audits of other schemes. The direction of the vector is clear even if the exact speed is uncertain.
In ten years, AI computers will cost as much as a smartphone
The horizon: verification, disclosure and the year that changed everything
Anthropic handled disclosure according to established norms: it notified the HAWK team in June, coordinated publication with NIST and with government and industry partners, and released open source code allowing anyone to reproduce the key recovery. It also collaborated with ETH Zurich, Tel Aviv University and the University of Haifa to launch CryptanalysisBench, a benchmark of 191 cipher-breaking tasks drawn mainly from NIST competitions.
But the question Anthropic itself left open in its statement is the one that defines what comes next: what do we do when a model finds a weakness in a cipher that is actually protecting systems in production? There is no established protocol for that situation. Responsible disclosure in cryptography assumes that the discoverer is a human team working on academic timescales. If the discoverer is a model working on API timescales, the whole framework needs rethinking.
A year out, language models have gone from being incapable of performing cryptanalysis even of the most basic ciphers to finding flaws in cryptographic designs that escaped years of expert review. The transition from finding implementation bugs to finding algorithmic weaknesses represents a qualitative leap. As the company wrote: "Many ciphers that protect modern systems have received less scrutiny than they deserve; they may still have important latent weaknesses that LLMs will soon be able to discover."