This Chinese AI Found 1,097 Critical Security Bugs Hiding for Decades — And Nobody Planned For It

I've been following AI developments closely for years, but this story stopped me cold. An open-source AI model from China just found over 1,000 critical security vulnerabilities in widely-used software — including bugs that have been sitting undetected for over 40 years. And the scariest part? The capability to do this wasn't even planned. It emerged on its own.

Meet GLM-5.3: The AI That Became a Security Expert by Accident

Z.ai, the Chinese AI lab behind the GLM model series, released GLM-5.3 — a massive 753 billion parameter mixture-of-experts model with a 1 million token context window. On the surface, it was built to be a state-of-the-art coding model. Z.ai fed it vulnerability-discovery data during post-training, expecting it to get better at reasoning about individual security bugs.

What actually happened was not what anyone planned.

Instead of just getting better at identifying single bugs, the model began forming coherent plans across complete exploitation chains. It wasn't just finding security holes — it was connecting them into full attack strategies. The capability compounded as training scaled, and nobody at Z.ai programmed it to do that. It emerged spontaneously.

The Numbers Are Staggering

Since GLM-5.2, Z.ai's models have surfaced 2,436 vulnerabilities across 269 open-source projects. Of those, 1,097 are rated critical or high severity — spanning kernels, browser engines, and network protocols. These aren't toy bugs in obscure hobby software. These are the kinds of vulnerabilities that, if exploited by a bad actor, could compromise enterprise systems, cloud infrastructure, and personal devices worldwide.

Here's the detail that really got me: the oldest bug in the set dates to approximately 1981. The average age of a discovered vulnerability is 26.6 years undetected. Security researchers have been auditing this code for decades and missed these flaws. An AI found them in what amounts to a few weeks of post-training compute.

Why Z.ai Delayed the Public Release

When Z.ai realized what they had on their hands, they made an unusual and responsible decision: they would not release the model weights to the public immediately. Instead, they're doing a two-week controlled rollout — distributing to security research partners first — while they work on safety controls.

The public Security Disclosure Ledger shows 53 CVEs already disclosed at launch, with 2,383 more still under embargo. That's 2,383 critical security vulnerabilities in widely-used software that haven't been publicly announced yet because they're still being patched by the projects that own the code.

The Dual-Use Problem Nobody Knows How to Solve

This story represents a genuine turning point for the AI industry. We've been talking about AI as a productivity tool, a creative assistant, a coding helper. GLM-5.3 shows that AI is becoming a security force multiplier — capable of doing in weeks what teams of human security researchers couldn't do in decades.

The dual-use nature of this is what makes it so complicated. The same model that finds 1,097 critical bugs for the defenders can — once the weights are released publicly — be used by anyone to find and exploit those same classes of vulnerabilities in software that hasn't been patched yet. Open-source AI models cannot be recalled once released.

Z.ai did the right thing by delaying. But what happens when the next version is even more capable, and a lab with less caution decides to just ship it immediately?

What Should You Do Right Now?

If you run any software infrastructure — whether personal servers, enterprise systems, or open-source projects — now is a good time to aggressively apply security patches. The CVEs being disclosed from this project are real, the embargo period won't last forever, and once those details are public, the race between defenders and attackers begins in earnest.

This is the kind of story that deserves far more mainstream attention than it's getting. An AI that accidentally became one of the best security researchers on the planet is either the best thing to happen to software security in decades — or a preview of what sophisticated attackers will soon be able to do. Probably both.

What's your experience? Drop a comment below! 👇 Does an AI that can find decades-old security bugs make you feel safer about our software infrastructure, or more worried about what bad actors will do with the same tools?