Claude Mythos finds weakness in test version of AES

Anthropic’s Claude Mythos Preview exposed vulnerabilities in a weakened version of the Advanced Encryption Standard, a widely used encryption system across the internet. The Reuters-backed research test highlighted how advanced AI models may begin probing the foundations of digital security, even though the weakness does not affect any cryptographic standard currently in use.

Anthropic claimed Mythos devised the cryptographic attack on AES almost entirely autonomously. According to a New York Post report, the chatbot spent nearly a week working on the problem before producing its approach, underscoring the pace at which AI systems are advancing in coding and cybersecurity tasks.

The US government adopted AES as its encryption standard in 2001, and it has long been considered among the strongest encryption systems. Traditional brute-force attacks and attempts to guess passwords or encryption keys remain practically incapable of defeating AES, but increasingly capable AI systems could create new pressure for stronger cryptographic safeguards.