TLDR Tech / InfoSec
How Harnesses and Post-Training Close the Open-Weight Bug Finding Gap
The author tested several open-weight models and Opus 4.7's ability to find the crackaddr vulnerability in four different variants. When using the Claude Code harness, only Opus 4.7 and GLM-5.1 consistently found all variants. However, other models performed better with the IronCurtain harness, which is specifically designed for security testing. GLM-5.1 performing significantly better than GLM-5 also speaks to the importance of post-training.
Read the full story