Claude finds cracks in HAWK and AES, and cryptography gets an AI-shaped cost problem

Cryptography has always been priced on the assumption that breaking it takes years of specialist labour. Anthropic's latest research suggests a model can now do serious damage for roughly $100,000 and a few days of compute, a shift that security teams and standards bodies will need to price into their threat models.

Anthropic said on 28 July that its Claude Mythos Preview model discovered new attacks on several cryptographic schemes. Most notably, it reduced the effective key size of HAWK, a post-quantum digital signature scheme, by 50%, using about 60 hours of work and roughly $100,000 in API costs. The HAWK authors were notified in June, ahead of publication.

On AES reduced to seven rounds, Claude Mythos produced an attack 200 to 800 times faster than the best previously known, after roughly three days of autonomous work and 1 billion output tokens, again costing about $100,000. Full-strength AES was not broken; reduced-round variants are the standard way cryptographers test a cipher's safety margin. The model also found a practical attack on LEA reduced to 13 rounds, needing fewer than 2^30 plaintexts and running in under an hour, and achieved full key recovery on Serpent-128 reduced to six rounds. Salsa20, Poseidon and SHA-1 saw only minor improvements, under 10 times.

Anthropic said the results were validated by human researchers before publication: two spent nearly a month, several hundred hours, learning cryptography to confirm the AES finding, while the HAWK result was checked by one researcher with a theoretical computer science background. Anthropic coordinated disclosure with NIST, the US government and industry partners. Anthropic's research 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."

The headline number for operators is the price tag: $100,000 buys a result that previously needed sustained expert attention, and that cost curve points one way. For firms building or auditing post-quantum systems, HAWK's halved effective key size is a direct signal to revisit assumptions before standardisation locks them in. Anthropic's coordinated disclosure with NIST shows it is treating this as infrastructure risk, not a research trophy, which raises the bar for any lab publishing offensive cryptanalysis without similar validation. Expect security vendors and standards bodies to start budgeting for AI-assisted red-teaming as a recurring line item, not a one-off exercise.