Anthropic's Claude Mythos Preview Identifies Weaknesses in Cryptographic Algorithms
Anthropic has released new research demonstrating that its Claude Mythos Preview model can identify vulnerabilities in cryptographic algorithms like HAWK and AES. Additionally, in collaboration with academic institutions, Anthropic introduced CryptanalysisBench, a benchmark designed to evaluate LLMs' capabilities in cryptanalysis. This milestone highlights the growing capability of large language models (LLMs) to assist in advanced security and vulnerability research, potentially transforming how cryptographic systems are audited. However, it also underscores the dual-use risks of AI, as these capabilities could be leveraged to compromise secure systems. The research includes full technical details of attacks on the HAWK and AES algorithms, along with the model's chain-of-thought reasoning for the AES attack. CryptanalysisBench consists of 191 tasks spanning six primitive families, including hash functions, block ciphers, and digital signatures.
## BACKGROUND
Cryptography is the practice of securing communication and data using mathematical algorithms like AES (Advanced Encryption Standard) to prevent unauthorized access. Cryptanalysis is the study of analyzing information systems to find hidden aspects, such as weaknesses or keys, to break cryptographic protocols. Benchmarks like CryptanalysisBench are crucial for systematically testing whether AI models possess the specialized reasoning required for these complex mathematical tasks.