Open-Weight AI Model Rivals Restricted Anthropic Tool at Building Browser Exploits

Anthropic's Frontier Red Team has found that Zhipu's openly available GLM-5.3 model can construct end-to-end exploits nearly as effectively as Claude Mythos Preview, a model the company restricted to critical partners due to security concerns. In a report published September 29, GLM-5.3 successfully built working exploits on 12 percent of tested scenarios, compared with 14 percent for Mythos Preview—while older models scored near zero.