Zhipu's GLM-5.3 ships a coding update and a 1,088-bug open-source audit

Zhipu released GLM-5.3 on Friday as a coding-focused update to its open-weight language model. The base is unchanged from GLM-5.2; the new work is in post-training. The release also ships a coordinated vulnerability-disclosure program: a 220-project, 1,088-medium-to-high-severity-bug audit run with Tsinghua, Nankai, and named Chinese security firms. The benchmark numbers are the hook. The audit pipeline is the move.

GLM-5.3 is not a new foundation model. Zhipu extends the GLM-5.2 base with additional post-training, per qbitai's coverage of the release. The vendor's own numbers, republished by NetEase Tech, put GLM-5.3 at 84.5% on the CyberGym white-box code-review benchmark, up from 77.2% on 5.2. The same tables name Anthropic's Claude as "Fable 5" at 83.8% and OpenAI's GPT as "GPT-5.6 Sol" at 83.6%. Those competitor labels are Zhipu's positioning language, not Anthropic's or OpenAI's official product names.

On Zhipu's ExploitBench, GLM-5.3 reportedly scores 54.4% versus 24.4% for 5.2. Zhipu's tables put Anthropic and OpenAI higher on that test (78.0% and 76.5%), so the win is narrow: white-box code review and certain exploit-discovery tasks, not the full agent leaderboard. The same release pushes Terminal-Bench 3.0 from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9, per qbitai. All of these are Zhipu-published numbers; no independent third-party reproduction is in the release packet.

The other half of the announcement is the audit. Since GLM-5.2 shipped, Zhipu ran a red-team exercise with Tsinghua's NISL lab, Nankai, and security firms Yunqi Wuyang (云起无垠), NSFOCUS (绿盟科技), Cyber Kunlun (赛博昆仑), DARKNAVY, QI-ANXIN (奇安信), and Tencent's Xuanwu Lab (腾讯玄武), per NetEase. Across 220 open-source projects, the program found 1,088 medium-to-high-severity vulnerabilities spanning system kernels, operating systems, browser engines, and network protocols. A separate count from qbitai reports 2,404 total vulnerabilities; the lower, severity-filtered figure is the more conservative read.

The most arresting finding is a critical bug in the DNS protocol dating to 1983, roughly 40 years old. The discovery was human-assisted: researchers at Tsinghua NISL and Yunqi Wuyang used GLM models as a tool to find and reproduce the issue, with GLM-5.3 reportedly used to confirm the underlying risk, per NetEase. A Type0 reader should not read this as the model finding the bug on its own. The model is one component in a human-led audit chain, and the disclosure predates the GLM-5.3 launch.

Zhipu is also announcing a program it calls "Open Source Shield" (开源的盾): free model quota for security audits on key open-source projects, plus new code-audit features in its coding tools. The labs and firms named in the release are concrete. None of the audit partners have published independent benchmark numbers; the CyberGym and ExploitBench scores above are Zhipu's own.

Day-one availability covers Zhipu's ZCode, AutoClaw, and GLM Coding Plan; the API is described as upcoming. Full model weights are scheduled to drop about two weeks after the release, according to NetEase Tech. That open-weights release is the second milestone to watch. Until independent labs reproduce the benchmark numbers and the audit partners publish their own write-ups, the most defensible read is narrower than the marketing: GLM-5.3 is a measurable post-training update with a real institutional partner set behind a 1,088-bug disclosure program, not a "Fable 5 killer" in the social-media sense. The next test is whether the open-weights drop lands on schedule and whether the DNS-class findings get a CVE.