Anthropic demonstrates that Claude Mythos Preview is capable of finding mathematical flaws in cryptographic algorithms

Anthropic has shown that its Claude Mythos Preview model was capable of discover real mathematical weaknesses in cryptographic algorithms, beyond simply searching for programming errors in the code.

The Artificial Intelligence (AI) laboratory published this Tuesday on its website all the details of this achievement achieved by simulating two attacks, one to HAWK, a post-quantum digital signature schemeand another to reduced AES, the most used symmetric encryption standard in the world.

The two achievements achieved by Claude Mythos Preview, the AI ​​that at the moment only several organizations on the planet have access to due to its repercussions on cybersecurity and biological and chemical threats, have been semi-autonomous (with minimal human guidance) in the attack on HAWK, and fully autonomous against a variant of AES encryption.

Anthropic explica the importance that cryptographic algorithms have in the daily lives of millions of peoplesince a simple visit to a website causes the user’s browser to communicate using an algorithm called a digital signature scheme.

In fact, the traffic generated between the person and the website is protected with symmetric encryption (code that allows the transmission of data that share an identical key). That is, without these systems, daily actions such as email, online banking and other internet uses would be easy to access for cybercriminals.

Regarding the effectiveness of Claude Mythos Preview in terms of cybersecurity, in the attack perpetrated on HAWK, Mythos not only improved the best known attack in just 60 hours of workbut instead discovered a previously unexploited mathematical symmetry and halved the effective strength of its keys.

The latter means that, to maintain HAWK’s level of security, the key size would have to be doubledwhich precisely sweeps away all the advantages that make HAWK an attractive proposition.

Regarding the mathematical symmetry discovered, since the security of the HAWK digital signature depends on the difficulty in solving a mathematical problem known as the Lattice Isomorphism Problemand since previous work had shown that finding a symmetry would allow the scheme to be attacked, Mythos Preview found one that was symmetry and that no one had detected.

Likewise, when Claude Mythos Preview faced reduced AES (seven of the AES-128’s ten rounds), he was able to almost completely autonomously design an improved technique called ‘Möbius Bridge’ to attack a reduced version of AES-128. What it achieved was an attack between 200 and 800 times faster than the best method known to date.

In this attack, Claude Mythos Preview was boosted at the beginning, since, according to the AI ​​laboratory, the model refused to commit it, maintaining that it was impossible to improve AES cryptanalysis. Ultimately, the process took him about three days of freelance work.

Anthropic has pointed out some details to take into account. Neither of the two attacks have been carried out on production systems (HAWK has not yet been deployed as it is only a candidate).

The cost of each result has been about $100,000 in APIand Anthropic notified the creators of HAWK prior to publication and coordinated disclosure with the US National Institute of Standards and Technology (NIST) in a responsible manner.

Anthropic, in collaboration with researchers from ETH Zurich and the universities of Tel Aviv and Haifa, created a ‘benchmark’ (CryptanalysisBench) to evaluate and further measure the ability of AI models to measure and further evaluate encryption algorithms in the future.