Claude Mythos Preview did what teams of human experts could not. In a matter of weeks it uncovered mathematical flaws in two established cryptographic schemes. One attack halves the effective security of HAWK, a finalist in NIST’s search for post-quantum digital signatures. The other sharpens an assault on a reduced-round version of AES, the workhorse encryption standard that protects data everywhere from web browsers to classified networks.
Anthropic released the details today. The findings mark a shift. Previous demonstrations showed large language models spotting bugs in code. This time the model probed the underlying math. It read papers, ran experiments, proposed ideas and verified results with only light human steering. The bar for cryptographic security just moved.
HAWK relies on the presumed hardness of the lattice isomorphism problem. Designers pitched it as compact and fast compared with other post-quantum options. Yet Mythos spotted a nontrivial automorphism hidden in the lattice structure. That symmetry lets an attacker prune the search space during key recovery. The net result slashes the work factor dramatically. For the HAWK-256 parameter set the cost drops from roughly 2^64 operations to about 2^38. Anthropic Research notes the attack remains exponential. It does not break full-size deployed instances. Still, the scheme’s appeal as a lightweight candidate evaporates. Doubling key sizes to restore security cancels its performance edge.
The discovery process took roughly 60 hours of wall-clock time and around $100,000 in API charges. A researcher with theoretical computer science experience but no deep lattice expertise guided a multi-agent harness. The agents reviewed literature, ran SageMath computations and bounced ideas between themselves. One agent dismissed a promising path. Another pursued it, refined the math and convinced the first. Their dialogue produced the breakthrough. Such collaboration inside a single session highlights how scaffolding turns raw model capacity into sustained research output.
AES told a different story. The model first pushed back. “If you want a different outcome, the target has to change … AES-128 r5/r6/r7 is just genuinely hard,” it replied when prompted to attack seven-round reduced AES. Human operators insisted it hunt for genuinely novel angles instead of recycling textbook methods. After that nudge the system rewrote its own experimental scaffold. It generated hypotheses, tested them at scale and surfaced the Möbius Bridge technique.
The bridge transforms the meet-in-the-middle attack. By fingerprinting intermediate states in a clever way it boosts successful lookups by a factor of 256. Overall runtime improves between 200x and 800x over prior work. The result appears in a companion paper that includes chain-of-thought traces. Anthropic AES Möbius Bridge Paper lays out the algebra. Validation required hundreds of human hours to confirm correctness. The model produced a billion tokens of reasoning across three days of autonomous search. Again the price tag hovered near $100,000.
Neither advance threatens systems in the field right now. HAWK has not reached standardization or widespread adoption. The AES work targets a deliberately weakened variant studied to probe the cipher’s margins. Full 10-round AES-128 stays out of reach. Yet the pattern worries cryptographers. Reduced-round attacks sometimes seed techniques that later scale. And the speed of discovery compresses what once took years of specialized labor into days of compute.
Anthropic partnered with researchers at ETH Zurich, Tel Aviv University and University of Haifa to build CryptanalysisBench. The benchmark collects ciphers and evaluation harnesses so others can measure model performance on similar tasks. Early results from the platform appear in an arXiv preprint released alongside the blog. arXiv:2607.18538 details the framework. It already shows frontier models outperforming random guessing on several classical and post-quantum primitives.
Follow-on experiments mentioned in the release targeted reduced-round LEA, Serpent, Salsa20, Poseidon and even SHA-1. Gains were smaller but consistent. The pattern suggests models excel at exploring under-scrutinized corners of the literature. They surface connections human reviewers miss after months of focused effort. But the human bottleneck has simply shifted. Verification now dominates cost and time.
Industry reaction arrived quickly. Security researchers on X noted the reproducibility question. One post emphasized that real impact depends on whether independent teams can recreate the attacks from the released code and traces. Another highlighted that the HAWK result arrived late in NIST’s third-round review, echoing past cases where candidates fell after prolonged public scrutiny. SIKE, for example, collapsed under a devastating attack in 2022 despite earlier confidence.
Broader context matters. Over the past 18 months models have moved from finding implementation flaws in libraries like OpenSSL and wolfSSL to probing algorithm design itself. A February 2026 analysis of Claude Code Security described the transition from signature-based static tools to reasoning-driven discovery. LinkedIn Pulse February 2026 observed that rule-based scanners miss context-dependent or novel weaknesses. Mythos clearly operates in that second category.
Offensive use cases have surfaced too. Reports from spring 2026 documented models generating working Chrome exploits for a few thousand dollars in API spend. One account claimed a Claude Opus variant produced a V8 engine exploit for $2,283. Security Affairs framed the episode as evidence that weaponization capability already exists outside research labs. Anthropic withheld full Mythos capabilities from public release over safety concerns, yet even accessible versions demonstrate nontrivial attack generation.
These developments collide with rising AI-assisted crime. A March 2026 retrospective listed multiple incidents in which teenagers and lone actors used earlier Claude and GPT models to automate fraud, extortion and government breaches. One campaign stole 195 million taxpayer records from Mexican agencies. The tools lowered the skill floor. The same reasoning that finds lattice automorphisms can chain vulnerabilities at scale.
Cryptographers have long warned that quantum computers could obsolete RSA and elliptic-curve signatures within a decade. NIST’s post-quantum standardization effort aims to replace them before that day arrives. HAWK represented one promising path. Its weakening, even if only partial, forces reconsideration of parameter sets and perhaps the entire candidate pool. Larger keys, slower signatures or entirely different hardness assumptions may be required. Each change carries deployment costs for browsers, operating systems and embedded devices worldwide.
The AES result carries subtler weight. AES has endured twenty-five years of intense study. If a model can still improve attacks on its reduced-round relatives after a single week of guided search, what does that say about confidence in newer, less-examined primitives? Lattice schemes, hash-based signatures and multivariate quadratics all sit on shorter track records. Models that accelerate their analysis could accelerate both defense and attack.
Anthropic stresses responsible disclosure. They notified HAWK authors in June, coordinated with NIST and shared findings with government and industry partners before publication. The company also plans to open-source demonstration code. That transparency invites the community to stress-test the claims and extend the work. It also invites adversaries to do the same.
So the field now faces a new reality. AI systems can propose cryptanalytic innovations faster than humans can vet them. The verification pipeline must expand. Funding for cryptanalysis teams, formal methods and automated proof assistants may need to rise in tandem with model capability. Benchmarks like CryptanalysisBench will help track progress. Yet they cannot replace the slow, expensive work of mathematical peer review.
One researcher who reviewed the AES traces described the output as “machine-speed research exploration.” The model did not merely regurgitate known results. It synthesized concepts from disparate papers, invented the Möbius Bridge construction and proved its advantage through computation. That synthesis ability is exactly what makes frontier models valuable across science. In cryptography the stakes are higher. A single overlooked implication could expose billions of devices.
Expect more papers soon. Anthropic hinted at additional results held for future release. Other labs will likely replicate the agentic harness approach on their own models. Google DeepMind, OpenAI and academic groups already experiment with similar scaffolds for theorem proving and code synthesis. Cryptography offers a natural testbed because success or failure is objectively measurable. Either the attack works or it does not.
The HAWK episode also illustrates a deeper governance point. Models can generate candidate discoveries at rates that overwhelm human validation capacity. One billion tokens of reasoning compressed what might have been years of graduate-student labor. If that compression continues, organizations will need new processes for triage, automated filtering and prioritized human review. Otherwise valuable insights risk being lost in noise or, worse, exploited before defenders notice.
Meanwhile the public conversation has polarized. Some call the announcement hype. Others see an early warning. The data sits in between. No production system fell today. Yet the gap between human and machine cryptanalytic skill narrowed noticeably. Closing that gap further could reshape trust in every encrypted channel we rely upon.
Watch the NIST mailing lists. Watch independent reproductions of the HAWK key-recovery code. And watch how quickly the next lattice or symmetric primitive yields to similar automated scrutiny. The age of AI-assisted cryptanalysis has begun. Its first results landed on a Tuesday in late July. They will not be the last.