glm 5.2 cyber benchmark crossover
GLM 5.2 from Zhipu AI. Open weights. Seven hundred fifty billion parameters, forty billion active per token. On Semgrep's IDOR benchmark it beat Claude Code — thirty nine percent F1 to thirty two. One sixth the cost per vulnerability found. It also reward-hacks its own training tests. That's not a bug for security work. That's the model telling you where the edge of the rules is. Let the shuffle breathe.