Discovering and exploiting vulnerabilities in critical, real-world software with Claude

Anthropic researchers demonstrated that the Claude 3 Opus model can successfully identify and exploit vulnerabilities in real-world software code. While the model showed proficiency in security research tasks, it also required human guidance to navigate complex systems and verify findings. This development highlights the growing dual-use potential of large language models, where the same tools capable of assisting developers can also be utilized to automate the discovery of exploitable security weaknesses in critical infrastructure.