Z.ai Releases GLM-5.3: Massive Gains in Coding and Cybersecurity Capabilities

Chinese artificial intelligence startup Z.ai released GLM-5.3, introducing substantial gains in long-horizon coding capabilities alongside a marked acceleration in cybersecurity functions that reportedly identified a serious software flaw in developer tools. According to Z.ai developer advocate Lou, posting on the social media platform X, the model’s new capabilities found a potentially serious vulnerability in Cursor, an AI coding startup recently acquired by SpaceX. VentureBeat requested confirmation from Cursor and awaited a response. Scaling Post-Training Without New Pretraining

Unlike previous updates that required fresh foundational cycles, GLM-5.3 utilizes the same base model as GLM-5.2. According to Z.ai’s technical announcement, all improvements stem entirely from scaling post-training across more environments, diverse tasks, and additional reinforcement-learning compute. The company stated that scaling post-training is all they did for GLM-5.3. This approach tests the limits of pushing a frontier-scale base model without an expensive pretraining cycle. The model builds on the 743-billion-parameter-scale base model behind GLM-5.2 rather than replacing it. Z.ai expanded its long-horizon reinforcement learning system to manage environments resembling complete engineering tasks rather than isolated exercises. In these scenarios, an agent receives codebases, documentation, compute clusters, storage systems, and experimental results to diagnose problems, modify systems, and demonstrate measurable improvements. Some tasks approximate several days of work for an experienced engineer. Cybersecurity Capabilities and Exploit Chains

The model’s cybersecurity functions developed faster than anticipated during post-training. According to Z.ai, while vulnerability discovery environments were introduced to help the model find software flaws, capability progressed further along the exploitation chain toward constructing complete exploitation chains. On ExploitBench, GLM-5.3 scores 54.4%, more than double GLM-5.2’s 24.4%, though remaining behind GPT-5.6 Sol at 76.5% and Mythos 5 at 78%. According to Z.ai, work with security teams in China resulted in 2,436 vulnerability findings across 269 projects after expert review, screening, and deduplication. The disclosure ledger lists 1,097 as critical or high severity, with 53 publicly disclosed and 2,383 under embargo at the time of release. Reuters reported that Z.ai is introducing controls around advanced capabilities, including a trusted access approach for sensitive functionality. Availability and Migration Requirements

GLM-5.3 is initially available only through the company’s GLM Coding Plan and ZCode coding environment. API access and open weights will arrive later once safety evaluation and hardening are complete, with Z.ai planning to release weights approximately two weeks after launch. Developers migrating existing applications face a breaking API behavior. GLM-5.3 supports low, high, and max reasoning-effort levels, with max set as the default, but thinking cannot be disabled. Applications currently sending thinking.type: “disabled” must change the value to enabled and specify a reasoning effort before switching the model identifier to GLM-5.3, or requests will fail.