Anthropic's Frontier Red Team Finds China's GLM-5.3 Nearly Matches Claude Mythos in Cyber Exploit Tasks
Covered by 2 sources

Anthropic's Frontier Red Team evaluated AI models on binary exploitation tasks, with GLM-5.3 achieving 4% and Claude Mythos Preview 6% success rates in control flow hijacks. GLM-5.3 also demonstrated capability to bypass safeguards and discover new browser vulnerabilities, signaling rising cyber risks from advanced AI.

