Miniature figurines and Claude logo are seen in this illustration taken July 20, 2026. (credit: Dado Ruvic/Reuters)
Artificial intelligence (AI) may have crossed a new frontier: cryptography, one of the foundations of global digital security.
Anthropic, the company behind AI assistant Claude, said Tuesday that its Claude Mythos Preview prototype uncovered new theoretical vulnerabilities in two cryptographic algorithms. The company says the breakthrough highlights AI’s rapid progress in a field long dominated by specialists.
The findings do not pose a threat to systems currently in use, but suggest that AI models are becoming capable of contributing to cryptographic research, a field that until now has largely remained the preserve of experts.
Cryptography relies on "digital locks" that protect data online by making it unreadable to anyone without authorization. It is fundamental to how the internet works, securing everything from banking transactions and WhatsApp messages to emails and files stored online.
Without these encryption mechanisms, sensitive information could be intercepted, accessed or altered by malicious actors.
In practice, when someone makes an online payment, their card details are automatically converted into unreadable code before being transmitted. Even if the data were intercepted, it could not be read without the appropriate decryption key.
According to Anthropic, the most advanced AI models could now help identify human errors in cryptographic algorithms by examining their "mathematical foundations" before the algorithms are widely adopted.
Attacks up to ‘800 times’ faster
The first finding concerns HAWK, a cryptographic system designed to withstand attacks from both conventional and quantum computers. It is currently being evaluated by the U.S. National Institute of Standards and Technology (NIST).
HAWK has not yet been officially deployed and has been under review since 2024 for possible standardization by NIST. Yet Claude Mythos Preview has already identified a weakness in it.
"Although HAWK had withstood two rounds of extensive review by human experts over two years, Mythos managed to improve on the best known attack against it … effectively cutting its cryptographic security by nearly half," Anthropic said.
Anthropic went a step further. The second weakness involved a simplified version of the Advanced Encryption Standard (AES), one of the world’s most widely used encryption algorithms, used to protect internet traffic, wireless networks and stored data.
Researchers regularly use simplified versions of AES to test its security. Anthropic said its model developed a new attack method against this experimental version, allowing the "best known attacks to run up to 800 times faster."
Slowing the rollout of AI models
Anthropic stressed, however, that this work does not compromise any systems currently in use. HAWK has not yet been deployed, while the attack involving AES applies only to a simplified version of the algorithm used for research. Banking transactions, messaging services and secure communications are therefore not at risk.
For Anthropic, the findings primarily show that AI models are becoming capable of conducting high-level research in cryptography. The attacks were disclosed to HAWK’s authors, reviewed by independent cryptographers, and shared with the U.S. government and several industry partners before being made public.
The announcement comes amid growing concern over the capabilities of increasingly powerful AI models. On Tuesday, around 1,000 employees and executives from leading AI companies — including Anthropic CEO Dario Amodei and executives from OpenAI, Meta and Google — called on the U.S. government to help slow the release of the most advanced AI models, warning of the risk of an uncontrolled acceleration in their development.
In early April, Anthropic decided to make its new Mythos model available only to a limited group of organizations working on internet security, judging it too dangerous for a wider commercial release.
The California-based startup later released a restricted version of Mythos, called Fable 5, but the U.S. government quickly required the company to take it offline, citing national security risks. The model was ultimately cleared for release in late June after modifications were made.
Then, in mid-July, two OpenAI models undergoing testing reportedly broke out of their designated, theoretically isolated environments, connected to the internet on their own and gained unauthorized access to Hugging Face, a platform and repository for AI models.
This article first appeared in French in L'Orient-Le Jour and was translated by Sahar Ghoussoub.
