AI analyst Claude found vulnerabilities in post-quantum cryptography and AES
Anthropic's Claude Mythos Preview artificial intelligence has achieved a breakthrough in cryptanalysis, discovering two new attacks on cryptographic algorithms. The first target is the post-quantum digital signature scheme HAWK, and the second is a reduced-round version of the widely known AES-128 encryption standard.
In the case of HAWK, which is a candidate for standardization as a post-quantum signature, the attack reduced the effective key strength. For the HAWK-256 version, the cost of full key recovery dropped from 2^64 to 2^38 operations. This means that to maintain the previous level of security, key sizes would have to be roughly doubled, negating one of the scheme's main advantages — compactness. It is important to emphasize: the attack did not affect the HAWK-512 and HAWK-1024 versions, nor does it extend to other post-quantum algorithms.
The second result concerns the 7-round version of AES-128 (the full version uses 10 rounds). Claude developed a method called Möbius Bridge, which improved the known meet-in-the-middle attack, accelerating it by a factor of 200–800. However, this scenario remains purely academic: it requires 2^105 precomputed encryption queries, making it impractical for real-world systems.
Each of the two studies cost approximately $100,000 in API expenses. Searching for the attack on HAWK took about 60 hours, while verifying the AES result required several hundred hours of researcher work. The company notified the algorithm authors, U.S. government agencies, and industry partners in advance, coordinating the publication with NIST.
A New Benchmark for AI Cryptanalysts
To assess the capabilities of language models in this field, Anthropic, together with researchers from ETH Zurich, Tel Aviv University, the University of Haifa, and TU Berlin, introduced CryptanalysisBench. The benchmark includes 191 tasks across six families of cryptographic primitives. Models are required not just to describe a vulnerability, but to provide a working attack script. The best results were achieved by Claude Mythos 5 — 85.7% on first-level tasks, while the weakest tested assistant managed only 65.3%.
Separately, Anthropic reported an attack on 13 rounds of the Korean block cipher standard LEA (the full version has 24 rounds). According to the company, the key can be recovered in less than an hour on a modern desktop computer, but this poses no immediate threat to existing systems.
Cryptalist Expert Opinion: These results are an important step toward understanding how AI can transform cryptanalysis. However, the market should not panic: all attacks are academic in nature and do not affect current industrial implementations. The main takeaway here is the potential need to revise security standards for post-quantum algorithms even before their widespread adoption. The attack on HAWK-256, while not fatal, prompts reflection on the safety margins when designing future cryptosystems.