Z.ai has unveiled GLM-5.3: a breakthrough in programming and cybersecurity
Chinese startup Z.ai has released an updated language model, GLM-5.3, marking a significant step forward in programming, agentic scenarios, and vulnerability discovery. Unlike previous iterations, the performance gains here were achieved exclusively through post-training, without changing the underlying architecture. This is an intriguing approach that demonstrates the potential of existing models is far from exhausted.
On the internal Code Bench benchmark, designed to evaluate coding skills, GLM-5.3 showed a 50% improvement over its predecessor. However, the most impressive results are seen in specialized tests. For example, in Terminal Bench 3.0, the new model scored 28.3 points compared to a meager 4.6 for GLM-5.2. In DeepSWE v1.1, progress reached 66.9 versus 46.2, and in Agents' Last Exam — 28.5 versus 23.8.
Particularly noteworthy are the results in the security domain. In the CyberGym test, the model achieved an 84.5% success rate, and in ExploitBench — 54.4%, more than double the previous version's result (24.4%). Within a two-hour budget, GLM-5.3 completed 105 tasks, whereas GLM-5.2 managed only 29. When the budget was increased to six hours, the gap persisted: 130 versus 39 tasks, respectively.
Real-world testing and public registry
Z.ai also conducted testing on real codebases in collaboration with several Chinese security teams. After verification and deduplication, the system identified 2,436 vulnerabilities across 269 projects, including 1,097 issues of medium and high severity. Only 53 findings were publicly disclosed; the remaining 2,383 remain under embargo. Notably, the average "lifespan" of such vulnerabilities is estimated at 26.6 years, with the oldest detected dating back to 1981.
For transparency, the company launched the Z.ai Security Disclosure Ledger — a public registry that records status, affected projects, and severity levels. In parallel, GLM-5.3 has already been deployed for GLM Coding Plan subscribers, becoming the primary engine for agentic programming and long-running tasks. Requests to older models are automatically routed to the new version.
Open weights for the model will become available two weeks after a security assessment and additional hardening. This is a sensible approach, given the sensitivity of the tool.
My take: GLM-5.3 is not just an evolutionary update but a serious bid for leadership in AI agents for cybersecurity. However, I should note that in July, Anthropic accused Z.ai of unauthorized distillation of GLM-5.2 using Claude and GPT responses. Given such a sharp leap in ExploitBench, it is worth closely monitoring further proceedings, although there is no evidence of violations yet.