From January to July 27, 2026, 45,207 security flaws were logged in the National Vulnerability Database, the public catalog of vulnerabilities. In seven months, the number nearly matched double everything 2025 recorded in the entire year.
It's not that software got worse. It's that artificial intelligence has learned to read source code hunting for holes — and it does so at industrial scale, without getting tired.
The news sounds good. The problem is what comes after it, and almost no one is talking about it: discovery has gotten faster than the fix.
What changed in the hunt for vulnerabilities
Old security tools looked for a known pattern: a snippet of code that resembled an already-cataloged error. AI models do something different. They read the code, connect clues that seem unrelated to each other, and reason about how to combine them into an attack — work that, until 2025, required an expensive and scarce human specialist.
The 2026 results show the size of the leap:
- An agent running the Gemini model found a critical flaw in Chrome that went unnoticed for 13 years, after countless human reviews.
- Firefox had 271 vulnerabilities located with Claude Mythos, plus 22 relevant issues with Opus.
- China's Z.ai claims its models found 2,436 vulnerabilities in 269 projects, with 1,097 critical or high severity, in infrastructure software like Linux, WebKit, and FreeBSD. The oldest was in code written in 1981 — open for 45 years.
On August 14, 2026, Z.ai announced GLM-5.3, which scored 84.5 on CyberGym — the benchmark that measures finding and validating real vulnerabilities from code. It ranked ahead of Claude Mythos 5 (83.8) and GPT-5.6 Sol (83.6). It's the first time an open model has led a security benchmark against the major labs' closed models.
Skepticism is warranted: all of these numbers were reported by the vendor itself, measured on its own infrastructure, without independent verification. Treat them as an order-of-magnitude indication, not an official scoreboard.
The number no one talks about: finding got easy, fixing didn't
Here's the part that changes decisions. Look at what happened to the companies with the best security teams in the world:
- Chrome: fixed 1,072 vulnerabilities in just two versions, in June 2026. That's more than the 1,036 fixed across the previous 23 versions, over roughly two years. Google started publishing fixes twice a week.
- Oracle: the July quarterly update brought 1,449 fixes, a company record. Only 64 came from external researchers — the rest came from automated tools and AI, according to Oracle's own attribution.
- Microsoft: 206 vulnerabilities in June, 622 CVEs in July, including 59 critical and 3 zero-day flaws.
Detection has become cheap and automatic. Verifying, prioritizing, fixing, testing, and publishing remains human work, at human pace. The bottleneck has shifted: it's no longer a shortage of people who can find the problem, it's a shortage of people who can apply the fix.
Why this is your problem, not just Microsoft's
If you only use off-the-shelf software, the math is simple: always update, and quickly. The vendor does the heavy lifting and your part is not delaying the update.
Now, if your company has proprietary software — that custom-built ERP, the customer portal, the app someone built a few years ago — the math is different. No one on the outside is going to audit that code for you. And it was written back when finding a flaw required a specialist poring over it for weeks.
That's exactly the kind of code the new generation of tools scans in minutes. The capability that finds a flaw in 1981 software won't have any trouble with your system from 2019.
Here's the uncomfortable part: the same capability serves both sides. A model that finds a hole to protect also finds a hole to break in. That's why Z.ai held back the release of GLM-5.3's weights for about two weeks for a security assessment — the company's first delay explicitly driven by cyber risk. The company also admitted this capability wasn't part of the plan: it added data expecting to improve detection of isolated bugs, and the model, on its own, started reasoning about complete exploit chains.
In other words: the tool that scans your system costs as little as $12.60 a month and will soon run on anyone's machine. The asymmetry that used to protect obscure software — "nobody's going to waste time looking at mine" — is gone.
What to do, in the right order
The answer isn't panic or hiring an expensive audit. It's organizing the basics, which almost no one has in place:
- Know what you're running. List systems, versions, libraries, and who maintains each one. You can't fix what you don't know exists — and it's common to discover a live system nobody's claiming responsibility for.
- Find out which ones have no maintenance. A system with no defined owner is the first to fall. If the answer is "the guy who built it left the company," you've just found your priority.
- Treat updates as routine, not a project. With vendors publishing fixes twice a week, the quarterly maintenance window has become a museum piece.
- Run the same tools before they're run against you. AI-powered scanning is affordable today. Using it on your own code is the cheapest way to find what an attacker would find.
- Prioritize by exposure. A flaw in a system open to the internet is worth ten flaws in an internal system. Start with what the world can see.
Notice that four of the five items are about inventory and process, not technology. The bottleneck is human — and organization solves more than buying a tool does.
The honest summary
2026 will close out with double the known flaws of 2025, and that's not going to slow down. Software didn't get worse: it got visible. The flaws were already there, some for 45 years, and now there's a way to find them at scale.
For anyone running proprietary software, the question is no longer "will someone look at my code?". Someone will. The question now is whether it'll be you first.
Facts compiled from technology news coverage in August 2026 (Daily Journal, Axios), the National Vulnerability Database, and the vendors' own announcements. GLM-5.3 figures are self-reported by Z.ai and have not yet been independently verified.
Perguntas frequentes
Can artificial intelligence really find security flaws in software?
Yes, and at scale. Unlike older tools, which looked for already catalogued patterns, current models read source code, connect clues with no apparent relationship and reason about how to combine them into an attack. In 2026, an agent running the Gemini model found a critical flaw in Chrome that had gone unnoticed for 13 years, and Firefox had 271 vulnerabilities located with AI support.
How many vulnerabilities were discovered in 2026?
45,207 flaws were registered in the National Vulnerability Database between January and 27 July 2026 — in seven months, almost double what the whole of 2025 recorded. The acceleration is attributed directly to the use of AI in vulnerability discovery.
If AI finds more flaws, doesn't software become safer?
It becomes safer and more demanding at the same time. The problem is that discovery became faster than repair: finding turned automatic and cheap, but verifying, prioritising, fixing and shipping is still human work. Chrome fixed 1,072 vulnerabilities across two June 2026 releases — more than in the 23 releases of the previous two years — and now publishes fixes twice a week.
My company runs an old custom system. Should I be worried?
More than those who use only off-the-shelf software. A custom system has no vendor publishing fixes, and nobody outside audits that code for you. It was usually written when finding a flaw required a specialist for weeks — protection that no longer exists. A model that finds a hole in code from 1981 will have no trouble with a system from 2019.
Why did Z.ai delay the release of GLM-5.3?
Because the same capability that finds flaws to protect finds flaws to break in. The company held the model's weights for about two weeks for safety evaluation — its first delay explicitly motivated by cyber risk. Z.ai also said the capability was not planned: it expected to improve detection of isolated bugs and the model began reasoning about complete exploitation chains.
Where should I start protecting my company's systems?
With inventory, not tooling. List what is running, with versions and an owner for each system; identify what has no maintenance; treat updating as routine rather than a quarterly project; run an AI scan on your own code before someone runs one against you; and prioritise what is exposed to the internet. Four of the five steps are organisation, not technology.