AI is beginning to do more than find software vulnerabilities. Anthropic says its Claude Mythos Preview model has now contributed new techniques for analyzing the mathematics behind cryptographic systems.
The model developed an improved attack against the post-quantum signature candidate HAWK and accelerated an existing line of attack against a seven-round version of AES-128, according to Anthropic. Neither finding affects encryption or signature systems currently used in production.
The immediate security risk is minimal, but the experiments offer an early look at how frontier AI models could accelerate cryptographic research—and potentially expose weak designs before they become standards.
Two cryptographic breakthroughs
According to Anthropic, Claude Mythos Preview developed an improved attack against HAWK, a digital signature scheme being evaluated through NIST’s post-quantum cryptography process.
The model completed the research in approximately 60 hours. Anthropic said the result substantially reduced HAWK’s estimated security level, suggesting that some proposed parameter sizes may offer less protection than expected. The company disclosed the finding to HAWK’s authors in June before coordinating its public release through a NIST mailing list.
Claude also improved an attack against a seven-round version of AES-128, eliminating one of the guesses required by earlier techniques and making the attack approximately 200 to 800 times faster, according to Anthropic.
Standard AES-128 uses 10 rounds. Cryptographers intentionally study reduced-round versions to test how well a cipher withstands increasingly sophisticated attacks, so the result does not allow attackers to decrypt data protected by production AES-128.
Research, not a practical exploit
Anthropic stressed that neither result requires changes to today’s security systems. The company said HAWK is still under evaluation and has not been adopted, while reduced-round AES is intentionally studied by cryptographers to understand how attack techniques evolve before they threaten real-world encryption.
The company also said it followed responsible disclosure practices, consulted academic experts to validate the findings, and shared the research with U.S. government and industry partners before publication.
Alongside the research, Anthropic announced CryptanalysisBench, developed with researchers from ETH Zurich, Tel Aviv University, and TU Berlin, to help evaluate AI systems on cryptographic research tasks.
What this means for cybersecurity
The work highlights how quickly frontier AI models are advancing beyond finding software bugs into mathematical cryptography research.
While there is no evidence that current AI models can break the encryption protecting online banking, messaging, or internet traffic, the research suggests they could increasingly help researchers uncover weaknesses in cryptographic designs before those systems are widely deployed. That could accelerate the review of future encryption standards, particularly as governments prepare for a future in which quantum computers may eventually threaten today’s public-key cryptography.
Also read: AI Cybercrime Report Warns of Emerging AI-Powered Threats