55: GLM-5.3: Frontier Coding with Emergent Cyber Capabilities
Show Notes
GLM-5.3: Frontier Coding with Emergent Cyber Capabilities
Episode 0055 | August 16, 2026
Summary
Zhiyuan (Z.ai) and THUDM released GLM-5.3 on August 14, 2026. The same base model as GLM-5.2. Every gain comes from post-training alone — more environments, more compute, the same open-source stack. What the engineers built was a coding model. What emerged from one month of scaling was a system that more than doubled its exploitation capability and began reasoning across complete exploitation chains.
They published the words "cyber capability developed faster than we expected." They built a public ledger of 2,436 real-world vulnerabilities. And they announced open weights in two weeks.
Key Numbers
| Benchmark | GLM-5.2 | GLM-5.3 | Change |
|---|---|---|---|
| Terminal Bench 3.0 | 4.6 | 28.3 | +515% |
| DeepSWE v1.1 | 46.2 | 66.9 | +45% |
| Agents' Last Exam | 23.8 | 28.5 | +20% |
| CyberGym | 77.2% | 84.5% | +7.3pp (SOTA) |
| ExploitBench | 24.4% | 54.4% | +123% |
| ExploitGym (2hr) | 29 tasks | 105 tasks | +262% |
Real-world findings: 2,436 vulnerabilities · 269 OSS projects · 1,097 critical/high · 2,383 under embargo · Oldest flaw from 1981 · Average age 26.6 years
The Emergence Story
Z.ai introduced vulnerability discovery data and environments into the post-training mix expecting improvement at finding and reasoning about vulnerabilities. What they did not expect was the speed of development further up the exploitation chain — from vulnerability identification through to multi-stage exploitation planning.
From the announcement: "GLM-5.3 did not simply become better at identifying isolated flaws: it began to reason across multiple stages of exploitation, forming coherent plans for complete exploitation chains."
The critical pattern: the further up the exploitation chain a benchmark sits, the larger the gain from GLM-5.2. Capability grew fastest exactly where they were furthest behind.
The Z.ai Security Disclosure Ledger
Z.ai published a public record of all findings at cvd.z.ai. The ledger is continuously updated as vulnerabilities move through the disclosure process, distinguishing publicly disclosed issues from those still under embargo.
Current ledger status (as of release): - 2,436 total findings tracked - 53 publicly disclosed - 2,383 under embargo - 1,097 critical and high severity - 107 critical · 990 high · 1,286 medium · 53 low - 45 years of collective impact - Oldest flaw: introduced in 1981
Training Stack
GLM-5.3 runs on the open-source post-training infrastructure built for GLM-5.2:
- slime — Open-source post-training framework for RL scaling. Megatron on training side, SGLang on rollout. Single dataflow keeps training, rollout, and data buffer unified so new environments plug in as data generation, not training loop changes.
- SAO — RL algorithm for long-horizon tasks with compaction. Keeps gains from collapsing on extended trajectories.
- IndexShare — Long-context processing infrastructure.
System-level optimizations for GLM-5.3 achieved 2.3× end-to-end RL training throughput improvement, enabling training over longer trajectories and more complex environments.
API Changes
Thinking is now mandatory. The disabled option is no longer supported.
Three reasoning effort levels:
- low — light
- high — enhanced
- max — deep (recommended for coding tasks)
Migration note: Applications currently using thinking.type: "disabled" must switch to enabled with reasoning_effort: low before updating to glm-5.3. Requests will fail otherwise.
Open Weights
Model weights will be publicly available two weeks after the August 14 launch, following safety evaluation and hardening.
Links
- Z.ai Blog: GLM-5.3 Announcement
- Z.ai Security Disclosure Ledger
- slime on GitHub (THUDM)
- SAO Paper (arXiv)
- IndexShare Paper (arXiv)
- ZCode — Try GLM-5.3
- Z.ai API Docs
- CyberGym Benchmark
- ExploitBench
DTF:FTL — Dispatches from the edge.