Yesterday, I ran an experiment, making further updates and iterations based solely on https://(github.com)/slowmist/openclaw-security-practice-guide. From v2.7 (current version) -> v2.8 (not yet released), the results were so-so. Ran into quite a few issues and had to rack my brain to outsmart the latest models from Gemini/Kimi. Surprisingly, the MiniMax model performed exceptionally well (minimax/orbit-20260303)... Looking back, the security practice guide fed into OpenClaw was trained like this. Very cautious work—if not done properly, it could easily lead to unexpected issues. Based on GitHub traffic data, the number of users is around 5,000. Future iterative training still needs to be conducted in "fully loaded" security scenarios and within a few real OpenClaw runtime environments. Until we reach what we consider a very stable version, we won’t release further updates... Security work can only be done slowly and carefully