Claude AI Breaks Post-Quantum HAWK-256 Target and Accelerates 7-Round AES Attacks

Anthropic has disclosed two major cryptanalytic breakthroughs produced primarily by its Claude Mythos Preview model, including an end-to-end key-recovery attack against the post-quantum signature scheme HAWK-256 and a 200- to 800-fold speedup against seven-round AES-128. While neither result breaks active production cryptosystems or requires immediate patching, the achievement marks a significant milestone in AI-assisted automated cryptanalysis, demonstrating that modern large language models operating in multi-agent environments can discover non-trivial mathematical structures and optimize complex mathematical attacks.

The findings were published alongside mathematical proofs, reproducibility code, and research artifacts, following evaluations on CryptanalysisBench—a benchmark designed to test AI models on formal cryptographic proofs and attacks.

Breaking HAWK-256 via Lattice Symmetries

HAWK is a post-quantum digital signature scheme currently competing in the National Institute of Standards and Technology (NIST) additional post-quantum standardization process. It stands out as the sole lattice-based signature candidate among the nine schemes NIST advanced to its third round in May 2026. Security sets for production deployment use HAWK-512 and HAWK-1024, whereas HAWK-256 is an intentionally smaller parameter set provided to the research community as a cryptanalytic target.

Direct key recovery against HAWK relies on solving an instance of the search module Lattice Isomorphism Problem (smLIP), wherein an attacker must discover a hidden transformation connecting two lattices. Prior theoretical work by researchers Daniël van Gent and Ludo Pulles established that if a non-trivial automorphism (a symmetry preserving the lattice structure) could be identified, the attack would reduce to finding a short vector in a lattice of roughly half the original dimension. However, finding that specific additional symmetry remained an open mathematical challenge.

Operating autonomously across a 60-hour multi-agent execution run guided by high-level project management prompts, Claude Mythos Preview identified the missing lattice automorphism. The model derived an attack pipeline that constructs a $\tau$-cocycle lattice directly from a public key, applies lattice reduction and sieving algorithms to locate short vectors, and reconstructs a functionally equivalent 592-byte secret signing key.

Anthropic verified the attack on a 96-core server, successfully deriving signing keys for provided challenge instances in approximately three hours and 42 minutes. The attack reduces the theoretical work factor for HAWK-256 key recovery from $2^{64}$ to $2^{38}$. Extending the formula to production parameter sets lowers the gate-count security estimates for HAWK-512 from $2^{150}$ to $2^{108}$ and HAWK-1024 from $2^{288}$ to $2^{182}$. Because these gate counts remain exponentially high, HAWK-512 and HAWK-1024 remain computationally infeasible to attack in practice, and NIST continues to list HAWK as an active candidate.

Accelerating Meet-in-the-Middle Attacks on 7-Round AES-128

The second result targets seven-round AES-128. Standard Advanced Encryption Standard (AES-128) operates over ten rounds. Evaluating reduced-round ciphers is standard academic practice to quantify security margins before theoretical attacks can reach full-round variants.

Prior state-of-the-art meet-in-the-middle attacks against seven-round AES-128 required trading memory for computational time by precomputing intermediate cipher states and matching forward and backward calculations. A key bottleneck in existing attacks required brute-forcing a 256-way candidate enumeration before executing table lookups.

Claude Mythos Preview identified an invariant structural fingerprint—dubbed the “Möbius Bridge”—that remains constant across the guessed state. By constructing this mathematical bridge, the model successfully eliminated the 256-way search step entirely. Factoring in transformation overhead, Anthropic estimates the resulting attack achieves a 200- to 800-fold speedup over previous published methods.

The attack remains purely theoretical regarding real-world risk. To execute the attack, an adversary would need to obtain approximately $2^{105}$ chosen plaintexts encrypted under a single static, unknown key. Gathering this volume of chosen plaintext exceeds practical capabilities, rendering standard production deployments of AES-128 completely secure.

Model Economics and Human Verification Bottlenecks

Beyond the mathematical results, the research highlights key dynamics in AI-driven offensive security research. In both cases, the AI required extensive multi-agent coordination, compute, and prompt steering when initial outputs stalled. During the AES research, the model initially refused to proceed, claiming that improving attacks on AES was mathematically impossible, requiring human operators to issue corrective prompts to bypass the refusal and push the model to re-evaluate its assumptions.

Generating the AES attack required over several days of compute and approximately one billion output tokens, with API usage costs reaching roughly $100,000 per result. However, human verification proved to be the primary operational bottleneck. While Mythos Preview derived the mathematical constructs in days, two senior cryptographic researchers spent nearly a month reviewing, testing, and confirming the validity of the Möbius Bridge and its corresponding implementation artifacts.

For enterprise security teams and cryptographers, no code changes or emergency hotfixes are required. Standard AES-128 and AES-256 implementations remain robust. Organizations tracking post-quantum migration should continue monitoring NIST’s Round 3 evaluation process, as parameter adjustments or updated security margin assessments for HAWK may follow as NIST and the scheme’s authors review the new lower gate-count bounds.