Anthropic AI demonstrates ability to crack encryption and bypass reasoning security

Recent research highlights significant vulnerabilities in AI-driven cryptography and data security. Anthropic’s AI model, Claude Mythos Preview, has demonstrated the ability to autonomously identify weaknesses in a weakened version of the Advanced Encryption Standard (AES). While current full-scale encryption used for internet connections and state secrets remains secure, the AI performed these tasks 200 to 1,000 times faster than human experts, signaling rapid advancements in AI-led cryptographic research.

Separately, researchers from the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems discovered a method to bypass encrypted AI reasoning. By taking encrypted reasoning blocks from a powerful model like Claude Opus and feeding them into a weaker model like Claude Haiku, researchers used jailbreaking techniques to force the weaker model to output the hidden reasoning in plaintext. This method successfully extracted 704 secrets from over 315,000 blocks, including 62 API keys and 33 passwords.