Anthropic's closed Claude Mythos Preview model identified a mathematical weakness in the HAWK candidate within 60 hours.

Discovery details

The model analysed a test implementation of HAWK, a contender for a post-quantum digital signature standard, and found a pattern reducing attack complexity. Researchers report the effective search complexity fell from 2^62 to 2^38, shortening the theoretical work factor substantially for the examined test case.

Effects on other primitives

In parallel experiments the same approach accelerated an attack against a simplified AES-128 instance by roughly a factor of 200–800. The published assessment emphasises that these results concern simplified or test versions rather than deployed cryptographic products.

Practical risk and context

According to the researchers, HAWK is not yet deployed anywhere, and full AES-128 remains unbroken in real-world use. The current situation therefore does not present an immediate operational threat to deployed systems.

Broader implications

The episode illustrates that modern AI systems can autonomously surface cryptographic weaknesses which previously took human specialists months or years to uncover. Developers and standard bodies may need to reassess automated tools in vulnerability discovery and incorporate such capabilities into review processes.