Chinese AI developer Zhipu has launched GLM-5.3, a coding-focused model that the company says uncovered thousands of vulnerabilities across real-world software. Zhipu's own testing places the model close to Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol on vulnerability discovery, though it trails both significantly on exploitation.
Benchmarks: discovery yes, exploitation less so
On CyberGym, a benchmark measuring vulnerability identification and validation, GLM-5.3 scored 84.5%, compared with 83.8% for Mythos 5 and 83.6% for GPT-5.6 Sol. But on ExploitBench, which tests how deeply a model can turn a flaw into a working attack, the gap widens sharply. GLM-5.3 scored 54.4%, while Mythos 5 hit 78% and GPT-5.6 Sol hit 76.5%.
"GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench," Zhipu said. "As we scaled post-training, this capability developed faster than we expected."
What the model found in real code
Zhipu said it worked with security teams in China to test the model against real-world codebases. After expert review and screening, the model identified 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues. The findings span system kernels, operating systems, browser engines, open-source infrastructure, and network protocols.
Zhipu's security disclosure ledger lists 107 critical and 990 high-severity findings. Fifty-three have been publicly disclosed, and 2,383 remain under embargo. The company also noted one unique vulnerability revealed in 1981, and that flaws in the dataset had existed in code for an average of 26.6 years before being identified. Zhipu did not say how many of the 2,436 findings were previously unknown or independently reproduced.
Why a coding model finds security weaknesses
Neil Shah, VP for research and partner at Counterpoint Research, said the result is an intentional side effect, not a technical exception. "We are commanding a stage where if we teach an AI to be a brilliant software engineer, you're accidentally teaching it how to be a good hacker, too," Shah said. "The exact same reasoning an AI uses to monitor code and fix bugs is what an attacker uses to find a weak spot and break through it."
The same reasoning appears in the model's training, which includes vulnerability-discovery data and realistic work environments. Zhipu also added environments that force the model to work against live compute clusters and storage systems, identify a flaw, implement an avoidance and deliver a measurable improvement.
Given the shared core, developers who work in this area often need a broader background in how AI-powered security works. A structured path for that sits in a separate resource: the AI Learning Path for Cybersecurity Analysts. Likewise, a developer staying current with these AI-driven security tasks may benefit from the AI Learning Path for Software Developers.
The bundle of vulnerability discovery, the question of disclosure, and open release creates pressure on the response window. Shah said the introduction of open weights raises the urgency: "If these AI-driven tools can discover thousands of flaws in real-world systems and anyone can directly access that capability, the next response window shrinks to near zero." Zhipu says it plans to release GLM-5.3's weights about two weeks after launch, after its own safety evaluation and hardening checks.
Your membership also unlocks:
