AI Found a Cryptography Flaw Two Years of Review Missed

Quick answer: On 28 July 2026 Anthropic published research saying an unreleased model, Claude Mythos Preview, found a previously unknown mathematical weakness in HAWK, a candidate in the NIST call for additional post-quantum digital signatures, after roughly 60 hours of work on a design that had already survived two years and two rounds of expert human review (Anthropic). The same model also improved a known attack on a deliberately weakened version of AES. Neither result breaks anything you use today, and no software needs to change (CSO Online).

Key takeaways

• Anthropic published the findings on 28 July 2026. The work was done semi-autonomously by Claude Mythos Preview using a scaffold built on Claude Code, with the human contribution described as mainly directing, organising and verifying.

• In HAWK, the model found a nontrivial automorphism, a hidden symmetry, in the lattice the scheme is built on. The attack is still exponential, so HAWK is not broken, but its effective security level is roughly halved, which would mean roughly doubling key sizes.

• For AES, it improved an existing meet-in-the-middle attack on a seven-round variant of AES-128, using a new fingerprinting method Anthropic calls the Mobius Bridge. Full ten-round AES-128 is untouched.

• Nothing deployed is affected. Anthropic notified HAWK's designers in June and coordinated public release with a NIST mailing list (CyberScoop).

• Verification, not discovery, was the expensive part. The AES result was found in about a week and then took two researchers close to a month to validate.

• We looked for a Nagpur or Vidarbha development connected to this story and found none, so there is nothing local to report here.

What Anthropic actually reported

The HAWK result

HAWK is a digital signature scheme, one of the remaining candidates in the NIST call for additional post-quantum digital signatures. Its security rests on the Lattice Isomorphism Problem, a question about whether two lattices are secretly the same object viewed from different angles.

Earlier academic work by Daniel van Gent and Ludo Pulles had shown that if a nontrivial automorphism existed in HAWK's lattice, key recovery would collapse to finding a short vector in a lattice of roughly half the original dimension. That paper opened the door but did not itself affect HAWK, because nobody had found such a symmetry. Mythos found one.

Precision matters here. The attack is still exponential in cost, so HAWK is not cracked. What changes is the safety margin: the scheme delivers roughly half the security level it advertises, and restoring that margin would mean roughly doubling key sizes. That is a serious problem for HAWK specifically, because compactness relative to rival candidates was the main reason to prefer it.

The AES result

AES-128 runs ten rounds. Cryptographers routinely attack cut-down versions with fewer rounds, because how far you can get into a reduced cipher is how the safety margin of the full one gets measured.

Working on a seven-round variant, the model devised a fingerprinting technique whose output does not change across the full range of guessed values, which removes an enumeration step that the previous best attack needed. Anthropic reports the improved attack as roughly 200 to 800 times faster, depending on how runtime is measured (The Hacker News). The AES you rely on when you open a banking app is unaffected.

What was not broken

This deserves stating flatly, because the headlines have not been careful. No deployed encryption was broken, no standard was withdrawn, and nobody needs to rotate a key or patch a library. What improved is the understanding of how much margin two designs really have.

Why this result is still worth your attention

Not because your data is at risk. Because of what it says about where machine capability now sits.

Cryptanalysis has an unusual shape as a research problem. The search space is enormous and mostly barren, so finding anything is very hard, but once a candidate attack exists the mathematics either holds or it does not. Problems that are expensive to search and cheap to check are exactly where a system that can generate and test enormous numbers of candidates has an edge.

Plenty of data work has some of that shape: feature search, model selection, query optimisation and hypothesis generation are all searches over large spaces where a proposal can be scored against data. The honest read for anyone in analytics is not that AI replaces the analyst, but that the generate-and-test half of the job is getting cheaper faster than the deciding half.

The human bill did not disappear

The detail most coverage skipped is the cost of being sure. The AES attack took about a week to find and close to a month for two researchers to validate.

That ratio is the point. When producing a plausible answer costs almost nothing, confirming it becomes the real constraint, and that cost is paid by people who understand the subject deeply enough to catch a wrong answer that looks right. Nobody in that story was doing prompt engineering. They were doing mathematics.

What this means in India

India has a specific version of this problem, and it is already measured. In July 2026 Nasscom published The State of AI-Native Talent in India, introducing an AI-Native Talent Index for early-career technology workers. The index assesses eleven dimensions, and alongside the expected ones such as prompting fluency and orchestration it explicitly scores judgment, cognitive independence and responsible use.

The reported picture is encouraging on adoption and cautious on depth: a large majority of early-career technology talent registers as AI-proficient or AI-native, while Nasscom's own commentary calls for stronger engineering fundamentals and independent problem solving so that the country builds AI-native engineers rather than merely AI-reliant ones (Business Standard).

The HAWK episode is a clean illustration of why that distinction is not academic. The scarce capability in that room was not the ability to get a model to produce an idea. It was the ability to tell whether the idea was correct. If you are entering a market where hiring is increasingly concentrated in AI and data roles, as our note on the AI job market in India sets out, that is the skill worth compounding. It also sits behind the national push described in India's sovereign AI model programme, which will need people who can evaluate models, not only run them.

What to do with this, practically

Learn to check, not only to generate. For every model output you accept, be able to say how you would falsify it: residual checks, holdout sets, sanity bounds and reconciliation against a second source, applied by habit rather than on request.

Go deep in at least one thing. The people who validated these attacks could do so because they knew lattice mathematics properly. The analytics equivalent is genuinely understanding statistics, or SQL semantics, or the mathematics behind the models you deploy. Our guide to essential skills for data scientists covers where to start.

Treat cryptography literacy as basic hygiene. You do not need to design ciphers, but knowing what hashing, signing and key rotation are for is part of handling data responsibly.

Do not over-read a single result. One model produced two findings in one specialised field, with heavy human scaffolding and verification. That is a milestone, not a forecast about your job.

The bottom line

The interesting thing here is not that an AI system embarrassed some cryptographers, because it did not. The process worked as intended: a candidate scheme was published, attacked, found wanting and reported before anyone deployed it. What is new is which participant found the weakness, and how fast.

For anyone building a data or AI career in India, the lesson is unglamorous and quite old. Value is moving towards the people who can judge whether an answer is right, which is a case for depth in the fundamentals rather than against it. Our PGP in Data Science and Analytics and Artificial Intelligence programmes are structured in that order, and how we research, source and correct articles like this one is set out in our editorial policy.

Sources and further reading

• Anthropic, Discovering cryptographic weaknesses with Claude, the primary research write-up published 28 July 2026.

• CSO Online and CyberScoop, on the scope of the findings, the disclosure process and the validation effort.

• The Hacker News, on the seven-round AES attack and the reported speedup range.

• Nasscom, The State of AI-Native Talent in India, and Business Standard's reporting on it.

FAQ

Frequently Asked Questions

Has AI broken encryption?

No. Anthropic and the security press reporting the result are both explicit that neither finding compromises production deployments or weakens any system in real-world use. The HAWK attack is on a scheme that is still a candidate rather than a deployed standard, and the AES result applies only to a deliberately weakened seven-round version of a cipher that ships with ten rounds.

What is HAWK and does anything use it today?

HAWK is a digital signature scheme and one of the remaining candidates in the NIST call for additional post-quantum digital signatures. Its security rests on the Lattice Isomorphism Problem. It is a proposal under evaluation, not a deployed standard, which is exactly why cryptanalysis of this kind happens before selection rather than after.

What is the Mobius Bridge technique?

It is the name Anthropic gave to a fingerprinting method the model devised while improving an existing meet-in-the-middle attack on a seven-round variant of AES-128. Because the fingerprint stays unchanged across the range of guessed values, an enumeration step in the older attack becomes unnecessary. Anthropic reports the resulting attack as roughly 200 to 800 times faster depending on how runtime is measured.

Does this mean AI can now do original research on its own?

It means an AI system produced a genuinely new result in a narrow domain where correctness can be checked. Anthropic describes the run as semi-autonomous, using a scaffold built on Claude Code, with the human contribution consisting mainly of directing, organising and verifying the work. Reporting notes that the AES result was found in about a week and then took two researchers close to a month to validate, so the verification effort was substantial.

Should this change how I study data science or AI in India?

Not the fundamentals. Python, SQL, statistics and the mathematics under your models remain the base. What the episode reinforces is that the valuable skill sitting on top is judgment: being able to tell whether an output is correct, and knowing how to check. The Nasscom AI-Native Talent Index published in July 2026 measures judgment and cognitive independence as explicit dimensions alongside prompting fluency, which points in the same direction.