AI Models Achieve Cyber Capabilities Breakthrough

A report from Anthropic’s Frontier Red Team reveals that advanced AI models, GLM-5.3 and Claude Mythos Preview, have demonstrated the ability to execute control flow hijacks in cybersecurity tasks. These models succeeded in 4% and 6% of trials respectively, a significant leap from earlier versions which failed entirely. This development marks a critical threshold in AI’s cyber capabilities, raising concerns about the spread of advanced hacking tools.