Anthropic revealed on July 28 that its unreleased Claude Mythos Preview model discovered previously unknown mathematical weaknesses in two cryptographic algorithms that human researchers had missed. The first finding targets HAWK, a post-quantum digital signature scheme that was a third-round candidate in the US NIST standardization process. Despite surviving two full rounds of expert human review over two years, Claude found a nontrivial automorphism in HAWK's underlying lattice that enables a faster key-enumeration attack, roughly halving the scheme's effective security — dropping a full key-recovery attack on HAWK-256 from approximately 2^64 operations to about 2^38.

The second result concerns AES, the most widely used symmetric cipher. Claude developed an improved meet-in-the-middle attack on 7-round reduced AES-128, running 200 to 800 times faster than the prior best-known attack. The attack required approximately one billion output tokens of compute, costing roughly $100,000 in API costs. However, AES-128 in real use has 10 rounds, so the result targets a deliberately weakened variant studied by researchers, not the encryption protecting real-world data.

Critically, Anthropic emphasized that neither result breaks deployed encryption and no production software needs to change. HAWK is only a candidate scheme and is not deployed, while the AES attack targets a reduced version. Independent cryptographers, including Johns Hopkins professor Matthew Green, validated the HAWK finding.

The discovery process itself is notable. For the HAWK attack, Claude worked semi-autonomously in a multi-agent harness with occasional human guidance from a researcher who was not an expert in lattice-based cryptography. The key idea was discovered by a pair of worker agents collaborating together — one prematurely rejected the idea, but the other found a way to exploit it, and after exchanging messages, both agreed on the attack. The entire discovery took about 60 hours.

For the AES attack, the process was almost entirely autonomous. A researcher built a scaffold allowing Claude to pose hypotheses, run experiments, and design attacks. Initially, Claude refused to engage, claiming improvement was impossible. After targeted prompting encouraging genuinely novel ideas, Claude rewrote its own agent harness and spent three days autonomously producing hundreds of millions of tokens before discovering a novel technique called a Möbius Bridge — a fingerprinting algorithm that eliminates one of the guesses an attacker needs to make.

The implications extend beyond the specific findings. If AI can find weaknesses that two years of expert human review missed, AI-assisted cryptanalysis should become a standard part of vetting future cryptographic standards. Anthropic partnered with ETH Zurich, Tel Aviv University, and TU Berlin to build CryptanalysisBench, a benchmark for evaluating LLM cryptanalytic capabilities. The company also disclosed additional preliminary results: a practical attack on 13-round LEA that runs in under an hour on a desktop, and a full key-recovery attack on 6-round Serpent-128.

Sources