An earlier version of this article was published on LinkedIn on July 29, 2026.
Anthropic’s recent work on AI-assisted cryptanalysis deserves attention—not because it proves that AI has “broken encryption,” but because it shows how quickly AI may compress the timeline between a new cryptographic design and a practical attack.
The most consequential result involved HAWK, a lattice-based digital signature scheme that had advanced to the third round of NIST’s additional post-quantum signature process. Anthropic reported that its Claude model, working with human researchers, found a practical attack against the HAWK-256 parameter set. The submission team subsequently withdrew HAWK from the NIST process.
That is a significant result. It is also easy to overstate.
What the HAWK attack changed
According to Anthropic’s research report, the attack reduced the expected cost of attacking HAWK-256 from roughly (2^{64}) operations to approximately (2^{38}). The attack does not apply to the other candidates in NIST’s process, nor does it represent a general break of lattice cryptography.
The result matters because the model did more than summarize existing research. It helped identify a new weakness, develop the attack, and turn it into something that human cryptographers could verify.
NIST’s Round 3 status page now records HAWK as withdrawn.
The AES result is not a break of deployed AES
Anthropic also described an improved attack on a reduced-round version of AES-128. Full AES-128 uses ten rounds. The research concerns seven rounds and therefore does not break production AES.
Even so, the result is interesting. The model improved a known attack that previously required about (2^{105}) chosen plaintexts and produced an implementation that ran roughly 200 to 800 times faster than previous work.
The boundary is important: progress against reduced-round AES helps cryptographers understand a construction’s security margin. It is not evidence that properly implemented, full-round AES-128 is currently broken.
The part I find most important
The headline result is the HAWK attack. The operational lesson is the changing ratio between discovery and response.
Anthropic reports that the HAWK work used about 60 hours of model time and cost roughly $100,000 in compute. Human verification took hundreds of hours. The reduced-round AES work required nearly a month of expert review.
That imbalance is worth watching. If AI systems can generate plausible attacks much faster than experts can validate them, the bottleneck moves. Discovery becomes cheaper, while verification, coordinated disclosure, standards decisions, product changes, and enterprise remediation remain human-intensive.
The long tail may be more exposed than the standards
Widely deployed algorithms receive sustained scrutiny. Many proprietary or lightly reviewed cryptographic constructions do not.
The most immediate risk may therefore be the long tail: custom protocols, vendor-specific schemes, legacy implementations, and cryptographic code that never received the attention given to AES or NIST candidates.
AI-assisted analysis could make it economical to inspect far more of that code. Some findings will be false leads. Others may expose weaknesses that have survived mainly because expert review was scarce and expensive.
What I would change now
Organizations do not need to panic or abandon proven cryptography. They do need to treat crypto-agility as an operational capability rather than a future architecture diagram.
That means knowing where algorithms, keys, certificates, protocols, and cryptographic libraries are used; understanding which systems depend on them; and being able to replace an algorithm or provider without redesigning an entire application.
It also means preserving the technical boundary of every new result. HAWK was a candidate-specific failure. Seven-round AES is not full AES. Accurate interpretation is part of security engineering.
The broader shift is still real. AI is beginning to change the economics and pace of cryptanalysis. The organizations best prepared for that shift will not be the ones that predict every future break. They will be the ones that can see their cryptographic dependencies, assess new evidence quickly, and change safely.