-
Anthropic says Claude Mythos Preview found a new attack that weakens the HAWK post-quantum signature scheme.
-
The AI also made an attack against a reduced version of AES 200 to 800 times faster.
-
Anthropic says neither discovery currently affects production security systems, but the findings show AI’s growing research abilities.
Anthropic’s Claude Mythos Preview has helped researchers find weaknesses in two cryptographic systems. The findings show how AI could help experts test security systems more deeply. Anthropic said the discoveries mark major progress in cryptography research. However, neither finding currently affects production systems.
The first discovery involved HAWK, a digital signature scheme built for a post-quantum future. Digital signatures help prove that digital information is real and has not been changed. Post-quantum systems aim to stay safe against future quantum computers.
According to Anthropic’s research report, experts had studied HAWK for two years. Mythos Preview then found a previously unknown attack in about 60 hours. The attack cut the security strength of HAWK by half.
However, Anthropic said HAWK remains a proposed system and has not entered production use. The Quantum Insider also reported that the finding affected a post-quantum signature candidate, not an encryption system protecting everyday internet traffic.
AI Finds New Attack Against HAWK
The HAWK discovery is important because researchers had already spent years reviewing the proposed system. Mythos Preview found a new way to attack its security design after roughly 60 hours of work.
Anthropic said the attack significantly reduced the security level that HAWK was expected to provide. The discovery highlights the value of cryptographic vulnerabilities—a hacker has recently claimed the sale of 10 crypto library zero-days for 200 Bitcoin. The discovery does not currently affect systems used in production because HAWK has not been deployed.
The finding, therefore, concerns a proposed security method rather than a system currently protecting people’s online data. Anthropic also shared the discovery with the people behind the algorithms, US government partners and industry partners.
The second finding involved AES, a widely studied system used to protect digital information. Mythos examined a reduced-round version of AES, rather than the complete version used in real-world systems.
Anthropic said the AI found a way to make an attack against the weaker AES version between 200 and 800 times faster. The model reportedly worked on the task for about one week.
CyberScoop reported that the research focused on a simplified version of AES. The work did not break the full AES standard.
Reduced AES Attack Does Not Break Full Encryption
The distinction matters because the research does not show that Mythos broke the AES system used to protect normal online activity. The finding does not mean that banking transactions, emails, passwords or ordinary web traffic protected by full AES are now exposed.
Anthropic said the result involved only a reduced version of the cipher. The company also said the HAWK finding has no current effect on production systems. HAWK remains a proposed system that has not been deployed.
The research still shows how AI can support complex security research. Anthropic said Mythos Preview completed most of the work with only occasional help from human researchers. Each discovery cost about $100,000 in API usage, according to Anthropic’s account of the research.
The company also disclosed the findings to relevant algorithm authors and government and industry partners.
Findings Show Growing Role of AI in Security Research
The discoveries come as researchers and organisations examine how advanced AI models could affect cybersecurity. The technology can help experts study security systems and search for weaknesses in complex mathematical designs.
Earlier in July, Reuters reported that Canada’s federal banking regulator warned major financial institutions about risks linked to advanced AI models, including Claude Mythos. The regulator’s concerns included the possibility that advanced AI could increase cyber threats. The technology could also reduce the time organisations have to find and fix security weaknesses.
Anthropic’s latest research does not show that today’s main encryption systems have been broken. Instead, the findings show that AI models can help researchers test cryptographic systems in new ways.
Mythos found a previously unknown attack against HAWK after years of expert review. It also improved an attack against a reduced version of AES. Anthropic’s results show that AI may become a useful tool for researchers who test security systems before attackers can exploit weaknesses.
The findings also show why experts continue to study the security of systems designed to protect digital information.
