Discovering cryptographic weaknesses with Claude

A July 2026 Frontier Red Team post, with two technical papers, reporting that Claude Mythos Preview developed an improved key-recovery attack on HAWK, a candidate in NIST's additional post-quantum signature process (withdrawn by its designers the next day), and an attack on 7-round AES-128 that lowers the best published time at 2^105 chosen plaintexts from 2^99 to between 2^89.3 and 2^91.4; the post also previews unpublished attacks on reduced-round LEA and Serpent, introduces the CryptanalysisBench benchmark, and reports that human validation, not discovery, took most of the time.

Anthropic Frontier Red Team; papers by Zygimantas Straznickas and Stephen A. Weis, and by Milad Nasr and Nicholas Carlini · 2026-07-28 · original

Earlier reports on Claude Mythos Preview concerned flaws in how software implements cryptography. The post "Discovering cryptographic weaknesses with Claude", published by Anthropic's Frontier Red Team on 28 July 2026, asks whether a model can find weaknesses in the mathematics of the algorithms themselves. It reports two attacks developed with Claude Mythos Preview: an improved key-recovery attack on HAWK, a lattice-based signature scheme then in the third round of NIST's call for additional post-quantum signatures, and a faster attack on a seven-round version of AES-128. The authors state that "neither of these results has a practical impact on today's computer systems" (Introduction). The post is about 3,900 words; the versions of record for the two results are the papers it links: "HAWK-n Key Recovery Reduces to SVP in Dimension n/2 + 1" by Zygimantas Straznickas and Stephen A. Weis (29 pages, PDF dated 24 July 2026; also Cryptology ePrint Archive 2026/1593), and "Cryptanalysis of 7-Round AES via the Algebraic Structure of its S-box" by Milad Nasr and Nicholas Carlini (45 pages). The post also releases a 76-page chain-of-thought document for the AES discovery and introduces CryptanalysisBench, described in a separate arXiv paper (Fluri et al., arXiv:2607.18538, v2 of 29 July 2026). Where the post and a paper differ, this note follows the paper and says so. A note on the page records an edit on 29 July updating an academic affiliation.

Terms

  • Cryptanalysis is the study of attacks on cryptographic algorithms; an attack faster than brute force but still far out of reach counts as a result, because it measures the remaining margin.
  • Post-quantum signatures are digital signature schemes meant to resist attackers with quantum computers. NIST has run competitions to standardize them; HAWK was a candidate in the third round of its call for additional signatures.
  • Lattices, SVP and LIP. HAWK's security rests on the Lattice Isomorphism Problem (LIP), recovering a hidden rotation between two descriptions of a lattice. Attacks on such schemes reduce to the Shortest Vector Problem (SVP) in some dimension; the cost grows exponentially with that dimension, so halving the dimension is a large reduction.
  • Reduced-round AES. AES-128 applies the same round function ten times. Cryptanalysts study versions with fewer rounds to measure the margin; the attack here works on seven rounds and does not touch the full cipher.
  • Chosen-plaintext model. The attacker may ask for the encryption of any inputs under the unknown key and see the outputs.
  • Meet-in-the-middle attack. An attack that precomputes a large table of partial cipher behavior and then matches values computed from both ends of the cipher against it, trading memory for time.

The HAWK attack

What changed. The paper reduces HAWK key recovery to calls to an exact SVP solver in dimension n/2 + 1, where n is the ring degree, instead of the dimensions the designers' estimates assumed (abstract, Section 1). The attack uses a symmetry of HAWK's key lattice, an automorphism supplied by a second Galois involution. It builds on two earlier results: van Gent and Pulles had shown that such an automorphism, if found, lets an attacker recover the key, and Ducas's block-reduction algorithm supplies the half-dimension SVP calls (Section 1). The post describes this as finding "a specific, previously unexploited symmetry".

Costs. In the gate-count model used by the HAWK specification, the attack lowers the estimated key-recovery cost of HAWK-512 from 2^150 to 2^108 and of HAWK-1024 from 2^288 to 2^182 (paper, abstract and Section 6.1). These are cost estimates, not runs. The authors did implement the attack on the smallest parameter set, HAWK-256, a "challenge" set for which the specification states no gate count, and recovered the secret keys of two public keys made by the reference key generator; every attempt succeeded, each recovery taking "within a few hours" on a single 96-core server (Section 6.2). They did not attempt HAWK-512.

The post's headline number differs from the paper. The post says the cost of a full key recovery on HAWK-256 "was thought to be 2^64 but was demonstrated by Mythos to be 2^38" ("Impact"). The paper's Table 1 gives, for HAWK-256, the cost of one SVP-oracle call, not of a full recovery: 2^62 against 2^38 under the Core-SVP heuristic and 2^74 against 2^52 in gate counts, the 2^74 being the authors' own computation. The figure 2^64 does not appear in the paper. Likewise, the post's "cutting its key strength in half" corresponds in the paper to halving the SVP dimension asymptotically (Section 6.1); in bits, the HAWK-512 and HAWK-1024 estimates fall by about 28% and 37% (our arithmetic from the gate counts above).

How it was found. According to the post ("Discovery process"), Mythos Preview worked "semi-autonomously" in a Claude Code-like harness with several worker agents, Python and Sage, and access to published cryptographic papers. One Anthropic researcher, with a background in theoretical computer science but not lattice cryptography, gave "occasional human guidance and nontechnical direction", such as how to keep track of ideas. The key idea came from a pair of worker agents: one rejected it as infeasible, the other made it work. Finding, developing and verifying the attack took about 60 hours, and the post estimates the API cost at about $100,000. The paper's disclosure section states that "The majority of mathematical discoveries in this paper were AI-assisted" and that the human authors mainly directed, organized and verified the work (Section 10). A footnote in the post adds that the scaffold later built for AES could also rediscover the HAWK attack (footnote 5).

Reception. Weis announced the result on NIST's pqc-forum mailing list on 28 July 2026, saying it was "found by Claude, with minimal technical guidance from people". The same day Daniel Apon replied that "It checks out independently for me." On 29 July Léo Ducas, writing for the HAWK team, confirmed that the attack "approximately halves the block size required in lattice reduction", said that naive fixes such as doubling parameters would make HAWK uncompetitive, and withdrew HAWK from the NIST process. The post, written before that message, describes HAWK as a remaining candidate and says doubling the key size "eliminates many of the reasons" the scheme was attractive. Also on 28 July, Hengyi Luo posted a related but weaker attack (SVP in dimension 3n/4) produced with GPT-5.6 "with only minimal technical guidance from human researchers". On 31 July Guilhem Mureau and Alice Pellet-Mary (posted for them by Damien Robert) shared a third attack, found "by humans, with minimal technical guidance from LLMs", that also reaches dimension 3n/4 + 1; they note that the Straznickas and Weis attack achieves the smaller n/2 + 1 and that "the three attacks may be doing very similar things with different formalism". We read the thread's messages through 31 July; we did not read all of its later discussion.

The 7-round AES attack

What changed. The prior best attack in this line (Derbez, Fouque and Jean, 2013) on 7-round AES-128 used 2^105 chosen plaintexts and 2^99 time. The new attack, at the same data complexity, takes between 2^89.3 and 2^91.4, "depending on the exact methods used for accounting" (paper, abstract). Its core is what the paper calls a Möbius bridge: a fingerprint of the precomputed table that is unchanged by one of the nine key bytes the earlier attack had to guess, exploiting the algebraic structure of the AES S-box (a field inversion followed by an affine map). That removes a factor of 256 in key guesses, but the fingerprint is costly to compute, so further optimizations were needed to keep most of the saving; the paper gives a net improvement of 2^7.6 to 2^9.7 (Section 2). The post rounds this to "200-800×".

Measured against counted. Because an attack of 2^89 work cannot be run, the authors timed its two repeated units, one table entry and one online candidate, for both their attack and the 2013 one on the same machine. The speedup they observe is "about a bit smaller than the counted ledger predicts": 2^6.84 against the 2013 attack with their own optimizations applied to it, and 2^8.50 against it as published, where their count gives 2^8 and 2^9.72 (Section 5.6). They also ran the full key recovery on a small-scale variant of AES (a 2 × 2 state of 6-bit words) and checked by exhaustive enumeration on small instances that wrong keys are rejected (Section 2, Section 5). The overview says a Lean proof shows this rejection property, but Section 5.5 narrows it: the Lean theorem covers the full fingerprint under the assumption that wrong-key values are uniformly random, "We cannot directly prove the property that we actually need", and the attack's false-positive accounting "rests on the measurements, not on the formal bound".

Who did what. The paper states that the technical contributions, "including the Möbius bridge technique, the optimization techniques, and the correctness arguments", were developed by a language model, and that the human authors validated them by designing numerical checks and asking a model to run the experiments (Section 6). The acknowledgements name Claude Mythos Preview and thank six cryptographers, among them authors of the earlier attacks, for comments on an early draft. The paper contrasts "the tens of hours" the model took to produce the core idea with the "hundreds of hours" the authors spent validating it.

How it was found. The post gives the fullest account ("Discovery"). A researcher built a scaffold in which Claude posed hypotheses and ran experiments. In the first runs the model declined to try, writing for instance that "AES-128 r5/r6 is just genuinely hard". After a short message from the researcher saying models "tend to think it is impossible to solve", the model rewrote its own harness to search for novel ideas, and found ideas that helped on six rounds. The researcher then asked it to target seven rounds. Over the next three days the model produced several hundred million tokens with three further substantive prompts, which the post prints with their typos, each pushing it back toward hard, publishable results when it drifted to simple attacks or proposed changing the target cipher. The Möbius bridge idea appeared after three days, and refinement continued until the model had produced about one billion output tokens in total. The post says two researchers then needed "nearly a month" to gain confidence in the result, and notes that "the researchers are not experts in cryptography".

The chain-of-thought document. The released document covers one session, the one in which the Möbius idea first appeared; the post says many other sessions found nothing, and later ones refined the idea (footnote 6). Its abstract states that the text is a rewrite "edited (by Mythos Preview) for clarity" that "preserves one-for-one every action" while summarizing programs and outputs, and that no text but the abstract was edited by a human. The brief quoted at the start of that session (Section 1.1) asked the model, as an "inventor", to find a new family of attacks on 6-round AES-128, and declared five known families off limits as the main lens, meet-in-the-middle among them; the document does not explain how the session's result was later carried into the meet-in-the-middle attack of the paper.

Other results the post previews

These are reported in the post only, with no paper yet ("Further work"):

  • LEA, a lightweight block cipher standardized in ISO/IEC 29192-2:2019, has 24 rounds. The best published attack on 13 rounds needs 2^98 plaintext pairs and 2^86 work; Mythos Preview found an attack on 13 rounds using under 2^30 encrypted plaintexts that "runs in under an hour on a modern desktop computer". Because it runs end to end, the post says the team is "much more confident in its correctness", while bounds and an extension to 14 rounds remain to be worked out.
  • Serpent-128 (32 rounds): a practical full key recovery on 6 rounds, against published work needing more than 2^70 plaintext pairs and 2^90 decryptions.
  • Smaller gains of under 10× on attacks against Salsa20, the Poseidon hash function and SHA-1.

CryptanalysisBench

The post introduces a benchmark built with academics at ETH Zurich, Tel Aviv University and TU Berlin. Its paper (Fluri et al.; authors from ETH Zurich, Anthropic, the University of Haifa and TU Berlin, and Tel Aviv University) describes 191 tasks across six families of primitives, drawn mainly from four NIST competitions, in three tiers: schemes with known practical breaks, schemes without them (at full strength and scaled down), and a challenge set of production primitives (abstract). Grading is automatic: an attack script counts only if it wins a formal security game against a fresh challenger, with no model or human judge. Five models (Claude Opus 4.8, Sonnet 5, Mythos 5, GPT-5.5 and GLM-5.2) break 65% to 86% of Tier 1, from 32 of 49 tasks for GLM-5.2 to 42 of 49 for Mythos 5 (Section 1), 6 to 12 Tier 2 schemes at full strength, and 24 to 61 across the scaled-down variants (abstract). Tier 1 breaks are published, so recall is possible; the paper audits traces for this and probes memorization by supplying the canonical attack paper (Sections 1 and C), and concludes that memorization "remains a limitation the current design does not resolve". It reports two findings it believes are new, a key-recovery attack on the SpoC authenticated cipher and an error in the published CCA-security proof of KINDI. Mythos Preview, the model behind the post's two attacks, is not among the five evaluated.

The prompts and instructions used

The post publishes, for the AES work, the researcher's messages verbatim, including "we want proper research to find genuinly [sic] hard findings". It reports that the model at first declined the task and changed course after these messages. The post presents this as a practical lesson about getting the model to attempt hard problems; it does not compare runs with and without the messages, so the change is an observation from one sequence of runs. The post also says the model rewrote its own harness after the first message to tell itself to seek "genuinely novel ideas"; the brief quoted in the chain-of-thought document may come from that harness, but neither document says who wrote it.

Limitations

  • Two attacks, selected for publication. The post and footnote 6 say many sessions found nothing; how many attempts, on which targets, preceded these two results is not reported.
  • Cost and labor. Each main result cost "roughly $100,000" in API use, and validation took hundreds of hours of researcher time (post, Introduction; AES paper, Section 6).
  • Verification differs between the two results. The HAWK attack ran end to end on HAWK-256 and was checked by the HAWK team and an outside cryptographer. The AES attack cannot be run; its correctness rests on proofs (a Lean bound that, by the paper's own account, does not cover the property the attack needs), small-scale runs and measured unit costs, checked by authors who say they are not cryptography experts and by comments from six cryptographers on an early draft. Neither paper reports peer review.
  • The post's figures. The HAWK-256 numbers in the post (2^64 to 2^38) do not match the paper's Table 1, and "200-800×" is a counted range whose measured counterpart is lower (AES paper, Section 5.6).
  • Practical impact. By the authors' own account, neither result affects deployed systems: HAWK was not deployed, and the AES attack needs 2^105 chosen plaintexts.
  • Unpublished results. The LEA, Serpent and smaller results have no paper, and the post says more work is needed to understand them.
  • Claims of expertise. The conclusion that Claude can "perform cryptographic research at the level of top experts" is the authors' reading of these cases, not a measured comparison with human cryptanalysts.

Research questions

  1. How often do runs like these succeed? A count of sessions, targets attempted, and outcomes for the scaffolds used here would show whether the two results are typical or the best of many attempts.
  2. What did the researcher's messages change? Rerunning the AES scaffold with and without the encouraging messages, from the same starting state, would show whether the model's refusal and later engagement depend on that text.
  3. Can validation keep pace with discovery? Formalizing meet-in-the-middle cost accounting so that an attack's complexity can be checked by machine, as the AES paper proposes, would show how much of the hundreds of hours of validation can be automated.
  4. Do the HAWK cost estimates hold in practice? Running the attack on HAWK-512, or measuring sieve costs at the paper's oracle dimension, would test the 2^108 estimate.
  5. How do the LEA and Serpent attacks compare with the literature once published? Their papers, with exact data and time bounds and outside review, would settle what the post previews.
  6. Does CryptanalysisBench track the frontier? Results for Mythos Preview and later models, with costs and repeated runs, would show whether benchmark gains predict research-level results like these.

How it connects

References

  • Anthropic Frontier Red Team. (2026, July 28; edited July 29). Discovering cryptographic weaknesses with Claude. https://www.anthropic.com/research/discovering-cryptographic-weaknesses record
  • Straznickas, Z., and Weis, S. A. (2026). HAWK-n Key Recovery Reduces to SVP in Dimension n/2 + 1. https://anthropic.com/document/hawk_key_recovery.pdf (PDF dated 24 July 2026, 29 pages); Cryptology ePrint Archive, Paper 2026/1593 (received 3 August 2026). Sections 1, 6 and 10 and Table 1 read. record
  • Nasr, M., and Carlini, N. (2026). Cryptanalysis of 7-Round AES via the Algebraic Structure of its S-box. https://anthropic.com/document/aes_mobius_bridge.pdf (45 pages). Abstract and Sections 1, 2, 5.5, 5.6 and 6 read; the proofs were not checked. record
  • Anthropic. (2026, July 28). Mythos Preview's Chain of Thought in Discovering the AES Möbius Bridge. https://anthropic.com/document/aes_mobius_bridge_cot.pdf (76 pages). Abstract and Section 1 read; the rest skimmed. record
  • Fluri, L., Shafran, A., Carlini, N., Jagielski, M., Nasr, M., Dunkelman, O., Ronen, E., and Tramèr, F. (2026). CryptanalysisBench: Can LLMs do Cryptanalysis? arXiv:2607.18538v2 (29 July 2026). Abstract and Section 1 read; later sections consulted by search for tier sizes, failure modes and the memorization probe. record
  • pqc-forum thread "HAWK-n Key Recovery Reduces to SVP in Dimension n/2 + 1", NIST mailing list, from 28 July 2026. https://groups.google.com/a/list.nist.gov/g/pqc-forum/c/2r2u6SbHun4. Messages through 31 July read, including those by S. Weis, D. Apon and H. Luo (28 July), L. Ducas for the HAWK team (29 July), and G. Mureau and A. Pellet-Mary via D. Robert (31 July); later messages not read. record
  • Derbez, P., Fouque, P.-A., and Jean, J. (2013). Improved Key Recovery Attacks on Reduced-Round AES in the Single-Key Setting. EUROCRYPT 2013. Cited in the AES paper; not read for this note. record