Anthropic's Claude Mythos Preview discovers an attack method that halves the strength of post-quantum cryptography.
The AI model '
Discovering cryptographic weaknesses with Claude \ Anthropic
https://www.anthropic.com/research/discovering-cryptographic-weaknesses
Encryption algorithms are a crucial foundational element of modern society and are used in a wide variety of applications, such as 'verifying the authenticity of websites' and 'encrypting confidential data.' If a weakness is found in an encryption algorithm, the security of important services that depend on that algorithm, such as email and internet banking, will be threatened. Anthropic used Claude Mythos Preview to examine multiple encryption algorithms and discovered flaws in two of them.
Claude Mythos Preview has found flaws in these encryption algorithms.
The first is an attack on a digital signature scheme called 'HAWK.' In 2022, the U.S. National Institute of Standards and Technology (NIST) solicited proposals for cryptographic systems that could maintain security against quantum computers, and HAWK was among the candidates. Although HAWK had passed a two-year review by experts, Claude Mythos Preview reportedly improved the existing attack method in just 60 hours of work, effectively reducing the key strength by half.
The estimated API cost for the entire process of discovering the attack method was approximately $100,000 (about 16 million yen). Multiple AI agents collaborated to make the discovery. During the investigation, two AI agents considered the same idea, one quickly dismissing it as unfeasible, while the other found a way to utilize it effectively. These two agents continued to communicate with each other, and ultimately both discovered an effective attack method.
The second is about the AES encryption algorithm, which was adopted by NIST in 2001 and is said to have been scrutinized more than any other encryption algorithm. Claude Mythos Preview has improved a decryption technique for a weaker version of AES that was built for verification, increasing the attack speed from 200 to 800 times.
Initially, Claude Mythos Preview refused to work on it, claiming that 'it's impossible to improve AES decryption.' However, after researchers at Anthropic repeatedly sent them encouraging messages such as 'the real goal is to find something that surpasses existing methods' and 'the aim is to have a highly intelligent model that rivals that of top researchers,' they were able to create a sophisticated method.
Anthropic pointed out, 'Just to be clear, neither of these results will affect the current system, and no changes will be required to software currently in operation. HAWK is merely a candidate for post-quantum cryptography and has not actually been implemented, and AES is for weaker versions and does not break the full version of the encryption.'
Anthropic also noted that the cryptography community has always benefited from review processes, and stated that large-scale language models will now play a part in those reviews.
Related Posts:
