AI Uncovers Decades-Old Cryptographic Weaknesses in Post-Quantum and AES Algorithms
Anthropic's Claude Mythos Preview AI has identified significant cryptographic weaknesses in the HAWK post-quantum signature scheme and a reduced-round AES version, flaws that eluded human experts for years.
Researchers at Anthropic have leveraged their advanced AI model, Claude Mythos Preview, to uncover fundamental mathematical flaws within prominent cryptographic algorithms. These newly discovered weaknesses affect the HAWK post-quantum digital signature scheme, a candidate in the NIST standardization process, and a reduced-round variant of AES, the ubiquitous symmetric encryption standard. While these findings do not pose an immediate threat to current production systems, they represent a significant advancement in AI's capability to perform adversarial review of cryptographic foundations.
The implications of such discoveries are profound, given that cryptographic algorithms underpin the security of nearly all digital interactions, from financial transactions to secure communications. Digital signatures authenticate online entities, while symmetric ciphers ensure the privacy of transmitted data. Historically, the identification of vulnerabilities in these core components has relied on human expertise and extensive peer review. Claude Mythos Preview's ability to identify algorithmic weaknesses, rather than just implementation bugs, marks a new era in security research.
The first major breakthrough targets HAWK, a scheme that had already undergone two years of rigorous human scrutiny as part of NIST's post-quantum cryptography competition. This competition aims to develop replacements for current public-key cryptography, which is vulnerable to future quantum computers. Claude Mythos Preview, after approximately 60 hours of semi-autonomous operation and an estimated API cost of $100,000, developed an improved key-recovery attack. This attack effectively halved the scheme's key strength by exploiting a previously unrecognized symmetry within HAWK's lattice structure, reducing the estimated cost of breaking HAWK-256 from 2^64 to 2^38 operations.
While the discovered attack remains exponential and impractical against larger key sizes, it significantly diminishes HAWK's attractiveness as a compact post-quantum solution. Anthropic has responsibly disclosed this finding to the HAWK authors and coordinated its release through NIST. The discovery highlights how AI can uncover subtle mathematical properties that human analysts might overlook, even after extensive review.
The second significant finding involves an enhanced cryptanalysis of a 7-round version of AES-128. Full AES-128, with its 10 rounds, remains secure against known attacks. By building upon existing meet-in-the-middle techniques, Mythos Preview developed a novel "Möbius Bridge" method. This technique eliminates a crucial guessing step, accelerating the attack by a factor of 200 to 800 compared to previous best-known methods for this reduced-round variant.
Anthropic has also reported promising preliminary results against reduced-round versions of other cryptographic algorithms, including LEA and Serpent, with smaller gains observed against Salsa20, Poseidon, and SHA-1. To facilitate further research and benchmarking, Anthropic has partnered with ETH Zurich, Tel Aviv University, and the University of Haifa to launch CryptanalysisBench, a platform for evaluating AI models in cryptanalysis.
These discoveries, while not requiring immediate changes to deployed software, underscore the growing potential of advanced AI in the field of cybersecurity. They demonstrate that AI can serve as a powerful tool for adversarial review, identifying vulnerabilities before they can be exploited in real-world systems. As AI capabilities continue to advance, human experts may increasingly shift their focus towards validating AI-generated research and developing defenses against AI-driven threats.
Anthropic is actively pursuing broader audits and plans to host academic workshops to explore the evolving role of language models in security research. The responsible development and deployment of such AI tools hold the promise of strengthening the cryptographic underpinnings of the internet and enhancing overall digital security for all users.
This new article provides further technical details on the HAWK-256 attack, specifying that the exploit leverages a previously unused symmetry within the scheme's lattice structure to achieve end-to-end key recovery. It also quantifies the AES-128 attack speedup, stating it removes a 256-way guessing step from a meet-in-the-middle attack, resulting in a 200 to 800 times acceleration.