Zhipu releases GLM-5.3 with an unplanned jump in cyber capability

Zhipu AI, trading as Z.ai, released GLM-5.3, an update that reused the same 743-billion-parameter base model as its predecessor GLM-5.2 rather than retraining from scratch, with the company attributing all capability gains to scaled-up post-training. Coding scores rose sharply on that basis: Z.ai reported Terminal-Bench 3.0 moving from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9, alongside a roughly 50% improvement on its own internal coding benchmark.

The more notable change was in cyber capability, which Z.ai said it had not specifically trained for and had not anticipated. According to Unite.AI’s account of the release, the company had added vulnerability-discovery data expecting isolated improvements in bug-finding, but the model instead showed compounding gains further up the exploitation chain, forming coherent plans across multi-step exploit development rather than just spotting individual flaws. GLM-5.3 scored 84.5% on the CyberGym vulnerability-discovery benchmark, up from 77.2% for GLM-5.2, which Z.ai reported as ahead of both OpenAI’s GPT-5.6 Sol and Anthropic’s Mythos 5 on that particular test, though Mythos 5 completed more tasks on the related ExploitGym benchmark.

Z.ai said it would withhold the open weights for about two weeks pending safety evaluation and hardening, releasing the model initially only through its own API, the GLM Coding Plan and its ZCode tooling. The gating followed a July assessment by the US Center for AI Standards and Innovation, which had already found that GLM-5.2’s safeguards would assist with agentic cyber-exploit development — making the delayed release of GLM-5.3’s weights a direct response to a capability regulators had already flagged as under-defended, in a model whose predecessor had shipped as the top-ranked open-weight release of its generation.