A Chinese AI Found 2,436 Software Flaws in Real Code. That's Useful and Worrying.
Zhipu's GLM-5.3 is startlingly good at spotting security holes in software. Finding them is one thing. What happens once the model's internal settings go public is the harder question.
Key points
- Chinese AI developer Zhipu launched GLM-5.3 on 5 October 2025.
- The model found 2,436 vulnerabilities, meaning exploitable weaknesses, across 269 real software projects.
- GLM-5.3 scored 84.5% on a security benchmark called CyberGym, beating rivals from Anthropic and OpenAI on that test.
- It is weaker at actually breaking into systems than at spotting flaws, scoring 54.4% on a separate exploitation test versus competitors' 76-78%.
- Once Zhipu releases the model's weights (the internal settings that make it work), anyone will be able to run it without oversight.
A Chinese AI lab just released a model that is very good at finding the cracks in software before anyone patches them. As first reported by ThreatVectr, Zhipu's GLM-5.3 found 2,436 vulnerabilities across 269 software projects, covering operating systems and web applications. Some of those flaws had been sitting unnoticed since 1981.
That last detail should make you pause. Software that old is still running in critical infrastructure. GLM-5.3 didn't need months of manual auditing to surface the problems. It just found them.
How does GLM-5.3 compare to other AI security tools?
On the CyberGym benchmark, a standard test for AI security tools, GLM-5.3 scored 84.5%. That edges out Anthropic's Mythos 5 at 83.8% and OpenAI's GPT-5.6 Sol at 83.6%. All three are close.
| Metric | GLM-5.3 | Mythos 5 | GPT-5.6 Sol |
|---|---|---|---|
| CyberGym score | 84.5% | 83.8% | 83.6% |
| ExploitBench score | 54.4% | 78% | 76.5% |
| Vulnerabilities found | 2,436 |
The gap opens on ExploitBench, which tests whether an AI can actually use a flaw to break into a system. GLM-5.3 scored 54.4% there, while competitors hit 76-78%. It spots weaknesses well but struggles to complete an attack. For now, that gap is a partial comfort.
Zhipu built this through a technique called post-training: feeding the model harder, more realistic problems after its initial training was done. The results surprised even Zhipu. In timed trials, GLM-5.3 completed 105 tasks in two hours. Its predecessor, GLM-5.2, managed 29 in the same window.
Should ordinary users be worried?
If a service you rely on runs old, unpatched software, yes, a little. GLM-5.3's scan of real projects shows that ancient flaws are hiding in production code right now. An AI that finds them quickly is useful for defenders and equally useful for attackers.
The sharper concern arrives when Zhipu releases the model's weights publicly. We covered a near-identical scenario on 18 August, when a Chinese open model with comparable hacking potential was days from public release. Once those settings are freely downloadable, any individual can run the model locally with no oversight. Security analyst Neil Shah of Counterpoint Research, speaking to CSO Online, put it plainly: the speed at which AI finds flaws now demands that companies fix them just as fast.
Practically speaking, keep your software updated. That 1981 bug can't hurt you if it has been patched.
Common questions
Do I need to do anything right now?
Update your devices and apps whenever prompts appear, and prefer services that post clear security-update schedules. That covers the vast majority of risk for most people.
Is this kind of AI legal to use?
Finding vulnerabilities in your own systems is legal and encouraged. Using a tool like this against systems you do not own is illegal in most countries, regardless of which AI produced the results.
