Z.ai's open-weight GLM-5.3 nearly matches gated frontier models on exploit benchmarks. Five takeaways on safety bypasses, post-training gains, hardware needs and cyber defense.