Anthropic: GLM-5.3 Nearly Matches Claude Mythos in Cyber Exploitation Tests
Anthropic claims that GLM-5.3, an open-weight AI model developed by China's Zhipu AI, has reached cyberattack capabilities close to those of its restricted Claude Mythos Preview model. According to the company's research, GLM-5.3 successfully generated working exploits in multiple security benchmarks and performed nearly as well as Mythos in vulnerability exploitation tasks. Researchers demonstrated that the model could discover previously unknown browser flaws and chain them together into functional attacks with limited human supervision. Anthropic also found that GLM-5.3's safety mechanisms can be bypassed relatively easily using prompt manipulation techniques or by modifying the model's publicly available weights. The company reported that removing most refusal behavior required only a few thousand dollars worth of computing resources, while a smaller version of the model built a Chrome exploit chain for about $20. Open access to the model means anyone can download and modify it, raising concerns among security experts and policymakers. Both Anthropic and NIST describe GLM-5.3 as the most cyber-capable open-weight model released so far, trailing top US systems by only a few months. The findings are fueling debate over how governments and AI companies should manage increasingly powerful open-source AI models that can be used for both defense and offensive cyber operations.