Chinese language AI Big Zhipu Unveils GLM-5.3, Edging Previous Western Rivals in Cybersecurity Benchmarks
Transfer over, primary coding bots, as a result of China’s synthetic intelligence scene simply dropped an enormous improve that’s sending shockwaves by the worldwide cybersecurity world.
Beijing-based AI innovator Zhipu AI formally launched its next-generation foundational mannequin, GLM-5.3, signaling a significant leap ahead from easy code technology to autonomous, deep-level digital protection. The discharge marks a vital shift for open-source AI, reworking digital assistants into energetic guardians of important digital infrastructure.
The mannequin’s cybersecurity chops proved eye-popping in worldwide benchmark testing. Within the CyberGym vulnerability reasoning analysis, GLM-5.3 posted a prime rating of 84.5 %, narrowly edging out Anthropic’s high-profile Mythos 5 mannequin, which scored 83.8 % when it debuted earlier this 12 months. On the ExploitBench check, designed to judge complicated reasoning and real-world vulnerability verification, GLM-5.3 scored 54.4 %, greater than doubling the efficiency of its previous-generation predecessor.
Zhipu didn’t simply depend on artificial lab exams to show its level. Working alongside researchers from Tsinghua College and Nankai College, in addition to top-tier Chinese language cybersecurity companies like Qi An Xin, Venustech, Tencent Xuanwu Lab, DarkNavy, CyberKunlun, and Cloudbreak, the crew put the AI by intensive red-teaming trials.
The sphere exams uncovered 2,436 verified vulnerabilities throughout 269 main software program tasks. Greater than 1,000 of these flaws have been categorised as medium-to-high danger, together with long-hidden bugs lurking inside core working techniques like Android, Home windows, and macOS, alongside broadly used cellular messaging apps. Most of the vulnerabilities had gone unnoticed by human engineers for years and even a long time.
Zhipu confirmed that the found flaws have been submitted to China’s official nationwide vulnerability databases, together with CNVD and CNNVD, initiating coordinated patching protocols to guard public and enterprise infrastructure earlier than malicious actors can exploit them.
