Claude Mythos Preview autonomously discovered thousands of zero-day vulnerabilities across major operating systems. The model identified vulnerabilities without prior CVE descriptions, surpassing GPT-4's performance in exploiting known flaws. The research from the University of Illinois showed GPT-4 exploited 87% of a 15-vulnerability dataset with CVE descriptions. Without descriptions, GPT-4 exploited only 7%. Anthropic's Claude Mythos Preview demonstrated a capability to find zero-day flaws independently. This undermines the industry's assumption that AI cannot discover unknown vulnerabilities.
Tap to vote and see what everyone thinks.
