Anthropic's Claude AI Achieves Breakthroughs in Cryptanalysis, Targeting Post-Quantum Schemes and AES

Anthropic has announced a significant advancement in cryptographic research, driven by its Claude AI model, specifically the Mythos Preview version. The AI has successfully derived an end-to-end key-recovery attack against HAWK-256, a post-quantum signature scheme. Post-quantum cryptography (PQC) refers to cryptographic algorithms designed to be secure against attacks from both classical and quantum computers. HAWK-256 is a specific type of lattice-based signature scheme, a category of PQC algorithms that rely on the difficulty of solving certain mathematical problems related to lattices. The attack developed by Claude AI exploits a previously unutilized symmetry inherent in the mathematical structure underlying the HAWK-256 scheme. This discovery is crucial as it probes the security of nascent PQC standards, which are vital for future data security in an era of increasingly powerful computing. The practical implications of this attack, even on a theoretical scheme, highlight the need for rigorous scrutiny of PQC candidates.

In parallel, Claude AI has demonstrated its prowess in cryptanalysis by achieving a substantial speedup for an attack targeting a reduced version of the Advanced Encryption Standard (AES). Specifically, the AI model has found a 200- to 800-fold speedup for an attack on seven rounds of AES-128. AES is a symmetric-key encryption algorithm that has served as a global standard for data security for over two decades, underpinning countless secure communications and data storage systems. While a seven-round attack does not represent a full break of the standard AES algorithm (which typically employs 10 or 12 rounds for AES-128), a significant acceleration of even partial attacks can reveal weaknesses in the algorithm's design or implementation. This finding underscores the ongoing need for robust cryptographic defenses and the continuous evaluation of established standards against evolving analytical techniques, including those powered by AI.

Anthropic has further contributed to the cybersecurity community by releasing an implementation of their findings related to the HAWK-256 attack. This implementation provides a verifiable benchmark, detailing an expected end-to-end runtime of approximately three hours and 42 minutes when executed on a 96-core server. This concrete performance metric allows security researchers to analyze the attack's feasibility, compare it with other cryptanalytic methods, and assess its potential impact on the security of similar lattice-based schemes. The proactive disclosure of such research by Anthropic aids in the collective effort to understand and mitigate emerging cryptographic threats, particularly as the world transitions towards quantum-resistant security solutions.