Anthropic's AI Cracks Encryption Standard Almost Fully Autonomously

Anthropic's Claude Mythos Preview, an AI model, has reportedly cracked a weakened version of the AES encryption standard, which was introduced in 2001 and considered nearly unbreakable. The attack was developed almost fully autonomously by the AI, with two human researchers spending nearly a month to verify the proposed method. According to the New York Times, the AI was 200 to 1,000 times faster than human crypto experts. Anthropic researcher Nicholas Carlini noted that as recently as 2025, AI models could not solve problems he solved at age 16, but now they conduct cutting-edge research. In addition to the AES attack, Mythos also orchestrated an improved attack on another cryptographic system called Hawk, which is designed to be secure against both traditional and quantum computer attacks and is a candidate for a new encryption standard by the National Institute of Standards and Technology. The Hawk developers confirmed that the attack method chosen by Mythos could work. While there is no acute danger to currently used encryption standards, experts warn of the rapid advancements in AI models in cryptography and potential long-term impacts on internet security, especially highlighting concerns about quantum computers and now AI models potentially threatening the mathematical core of internet security standards.