Models GLM 5.3-flash showed 84.5% in the cybersecurity benchmark CyberGym

Researchers tested the capabilities of GLM 5.3-flash in cybersecurity tasks, recording a result of 84.5% on the specialized CyberGym benchmark. The model can be run locally on consumer equipment, making it an accessible tool for automating vulnerability detection. Particular attention is paid to versions with removed security filters, so-called "abliterated" models. During testing, such configurations demonstrated the ability to bypass standard protection mechanisms, performing code analysis and system exploitation tasks without the limitations imposed by developers during training. The technical accessibility of GLM 5.3-flash allows it to be used for real-time automated search for attack vectors. The high accuracy of 84.5% on CyberGym indicates that current software protection methods may be ineffective against specialized AI agents operating without ethical constraints. Experts from the community note that the industry has about a year to adapt protection systems to new threats. The growth in performance of local models forces a reevaluation of security approaches, as controlling the use of AI tools outside of cloud APIs becomes virtually impossible.