XNews

GLM-5.3 Launch and Cybersecurity Coding Advancements

GLM-5.3 Launch and Cybersecurity Coding Advancements

Zhipu AI launched GLM-5.3 on Aug 14, 2026: same 743B base model as GLM-5.2, with post-training RL only. Official announcement cited top-tier coding and agentic capabilities plus a cybersecurity leap setting a new open-model standard. Coding: Terminal Bench 3.0 4.6->28.3; DeepSWE v1.1 46.2->66.9; Agents' Last Exam 23.8->28.5; Code Bench Max 34.5% at ~75K tokens vs 23.4% at 96K; Fable 5 Code Bench 39.5%, ExploitBench 78.0 vs 54.4; open-source first, global second tier. Cybersecurity: CyberGym 84.5% vs GPT-5.6 Sol 83.6%, Fable 5 83.8%; ExploitBench 24.4%->54.4%; ExploitGym 29/39->105/130; exploit-chain planning emerged from RL scaling. Real code: 2,436 vulnerabilities in 269 projects, 1,097 medium/high-risk, oldest ~40 years, 53 disclosed, 2,383 embargoed, Security Disclosure Ledger online. Stack: IndexShare, SAO, slime; more environments/tasks/compute; senior engineer multi-day scenarios; automated synthesis with human involvement. API live for GLM Coding Plan; supports ZCode, Claude Code, OpenCode; thinking levels low/high/max only; off-peak 50% credits; ZCode >98% cache and 1.5x bonus through August; weights public in two weeks.

Key insights

Open-source coding lead with closed-source gap

GLM-5.3 recorded the highest scores among open-weight models in multiple programming benchmarks. Terminal Bench 3.0 improved from 4.6 for GLM-5.2 to 28.3. DeepSWE v1.1 rose from 46.2 to 66.9, and Agents' Last Exam rose from 23.8 to 28.5. On Zhipu Code Bench Max level, GLM-5.3 reached 34.5% completion with about 75,000 tokens, compared with GLM-5.2 at 23.4% using 96,000 tokens. Against closed-source frontier models, gaps remain: Fable 5 reached 39.5% on Code Bench and 78.0 on ExploitBench, while GLM-5.3 scored 54.4 on ExploitBench. The overall position is open-source first and global second tier.

Cybersecurity benchmark leap

In cybersecurity evaluations, GLM-5.3 achieved a CyberGym white-box source audit score of 84.5%, the highest result, above GPT-5.6 Sol at 83.6% and Fable 5 at 83.8%. ExploitBench improved from 24.4% for GLM-5.2 to 54.4% for GLM-5.3, more than doubling. ExploitGym rose from 29 out of 39 completed tasks to 105 out of 130. The progress was larger in later stages of the vulnerability exploitation chain. Zhipu stated the model began autonomously planning complete exploit chains after post-training included vulnerability discovery data and environments.

Emergent exploit planning

Zhipu described the exploit-chain planning ability as 'emergent,' meaning it was not specifically trained but appeared while scaling reinforcement learning. The team expected improvements in vulnerability identification but observed autonomous planning of complete exploit chains. This claim, if accurate, is a relevant signal for AI safety because the capability emerged from expanded RL scale rather than direct training.

Real-world vulnerability discovery and disclosure

From GLM-5.2 onward, Zhipu worked with domestic security teams to scan real codebases. After expert review and deduplication, the model found 2,436 vulnerabilities across 269 open-source projects. Of these, 1,097 were medium or high risk. The projects included operating system kernels, browser engines, network protocols, and other core infrastructure. Some vulnerabilities had been present in code for decades, with the oldest dated around 40 years. As of the post, 53 vulnerabilities were publicly disclosed and 2,383 remained under embargo. Zhipu also published a Security Disclosure Ledger for tracking.

Public weight release contrasts with restricted vulnerability engines

This approach parallels Anthropic's April announcement that Claude Mythos found a large number of zero-day vulnerabilities, but Claude Mythos was strictly restricted by the U.S. government to a small number of vetted security organizations. In contrast, GLM-5.3 weights are scheduled for public release in two weeks. The post states that a vulnerability discovery engine with comparable capability is about to enter the public domain in open-source form.

Training stack and environment scaling

GLM-5.3 kept the same base model as GLM-5.2 and expanded post-training scale. The stack includes IndexShare for long-context handling, SAO for long-horizon RL tasks, and the open-source slime framework for large-scale asynchronous training. GLM-5.3 added more environments, more diverse tasks, and more training compute. New training environments simulate senior engineer work, with some tasks equivalent to several days of work for experienced engineers, such as diagnosing training bottlenecks and delivering end-to-end acceleration in an ML infrastructure environment. An automated environment synthesis pipeline uses research agents to generate executable long-horizon environments and evaluation agents to verify tasks, with human involvement still required.

API availability, pricing, and supported tools

The GLM-5.3 API is already online for all GLM Coding Plan users. It supports ZCode, Claude Code, OpenCode, and other coding agent tools. Thinking cannot be disabled; users choose low, high, or max levels, with max recommended for programming tasks. Billing moved to a credit system. Off-peak periods, all times outside weekdays 14:00-18:00 Beijing time, consume 50% of standard credits, including weekends. ZCode usage offers a claimed cache hit rate above 98% and a limited-time 1.5x quota bonus through the end of August, equivalent to 180% of standard quota.

Source posts