Claude AI finds new cryptographic weaknesses, Anthropic says

What happened: Anthropic said its Claude Mythos Preview model discovered new attacks against several cryptographic schemes, according to research published 28 July 2026.

The numbers: Against the post-quantum signature scheme HAWK, the model halved the effective key size after about 60 hours of work and roughly $100,000 in API cost. Against AES reduced to 7 rounds, it produced an attack 200 to 800 times faster than the best previously known, using 3 days of autonomous work, 1 billion output tokens and about $100,000.

The details: The model also broke LEA reduced to 13 rounds with a practical attack needing fewer than 2^30 plaintexts in under an hour, and achieved full key recovery on Serpent-128 reduced to 6 rounds. It made only small improvements, under 10 times, against Salsa20, Poseidon and SHA-1.

The catch: Full-strength AES was not broken. Reduced-round variants are how cryptographers measure a cipher's safety margin, not real-world vulnerabilities on their own.

Who's affected: Human researchers still had to check the work. Verifying the AES result took two researchers nearly a month, several hundred hours, during which they had to learn cryptography themselves. The HAWK result was checked by one researcher with a theoretical computer science background.

The context: The HAWK authors were notified in June. Disclosure of the findings was coordinated with NIST, the US government and industry partners.

In their words: Anthropic said: "Without secure cryptographic systems like these, your email, online banking, and other internet use would be open to cybercriminals, who could intercept or modify your communications."

Why it matters: The results show an AI model producing original findings in one of the most demanding fields of mathematics and security, rather than just applying known techniques.

What's next: Anthropic gave no timeline for further cryptographic testing beyond the disclosed coordination with NIST and industry partners.