The Vulnerability Remediation Problem: Project Glasswing and Patch the Planet are Finding Flaws Faster Than Teams Can Fix Them
In the first nine months of 2025, 35,448 vulnerabilities were published. Over the same stretch of 2026, that number hit 72,812, and the third quarter alone more than tripled.
There are two ways to read that. The hopeful one is that AI has finally given us tools good enough to find problems that were sitting in our software all along. The worried one is that it’s exposing weaknesses faster than anyone can close them. I’ve gone back and forth on this, so I’ll make both cases before telling you where I land.
One ground rule first. A CVE is a tracking ID for a publicly disclosed vulnerability. It counts disclosures, so doubling the count doesn’t double your problems. Still, every one of those disclosures ends up on somebody’s desk as a question: do we run this, and what do we need to do about it?
The Optimist’s Case (and Its Catch)
Software has always had vulnerabilities. There’s an enormous amount of them, and human researchers only have so many hours. AI changes that math, and a long list of major companies is behind the effort. Anthropic’s Project Glasswing brought together Microsoft, Google, Apple, Cisco, the Linux Foundation, and others to point advanced models at critical software. OpenAI’s Patch the Planet pairs AI-assisted research with human engineers who validate findings, build and test patches, and work alongside open-source maintainers.
Those human engineers carry the optimistic case. A vulnerability report on its own just creates work for somebody, while a reliable fix in a shared component can protect many products at once. That’s why it matters that OpenAI’s Linux research produced candidate patches alongside the 1,724 vulnerabilities it reproduced.
The catch shows up in Glasswing’s own numbers. A month in, Anthropic reported that 75 of the 530 high- or critical-severity bugs it had disclosed were patched, and some maintainers had asked it to slow down so they could keep up. By October, it had more than 29,000 candidate findings, about 6,000 reviewed by humans, and said it was bottlenecked on human capacity to validate the rest.
I’d take that as one of the more honest data points in this debate. Teams with some of the most capable models available are hitting the same wall everyone else hits, which is that discovery scales faster than validation and repair, at least for now.
Attackers Don’t Have a Change Approval Process
While defenders work through the backlog, attackers are putting the same kind of capability straight into operations.
ARTEX is a good example. It’s an open-source, AI-driven penetration testing framework. You give it an objective and it coordinates agents and tools to pursue it. It was released in late July, and by early October CrowdStrike had tied it to a malicious campaign.
Then it showed up much closer to home. Microsoft attributed a mid-October campaign against ConnectWise Cloud Backup customers to Storm-0940, a group it assesses as China-state-aligned. Using ARTEX, the attackers abused compromised ConnectWise application identities to sign in to Microsoft 365 tenants that had already authorized those apps. Microsoft counted more than 2,250 tenants targeted at up to 2,900 sign-in attempts per minute, with some reaching privileged access in under three hours.
That campaign didn’t depend on an unpatched vulnerability, since the attackers exploited access customers had already granted. It’s a clear look at how AI speeds up an attacker’s clock. Compare that with how long your last emergency change took to get approved.
Exploitation of vulnerabilities is speeding up, too. Mandiant’s M-Trends 2026 puts the mean time to exploit at an estimated -7 days, meaning exploitation is routinely starting before a patch even exists. Zero-days drag that average below zero, so it doesn’t mean every bug gets hit early, but the trend is hard to argue with.
Once a patch ships, it also doubles as a map. Attackers can compare old and new code, spot the security change, and work backward to the flaw, a technique called patch diffing. In a controlled Anthropic study using patches that had been public for at least 90 days, a model built a working exploit against Firefox’s JavaScript engine in under an hour.
Where the Race Is Lost
Publishing a fix starts a race, and the defender’s side of it has a lot more steps. At Galactic, we describe them as Find, Fix, Verify, Repeat. Each one hides work, dependencies, and decisions, and speeding up one doesn’t speed up the others.
Find: Find is two problems in one. A researcher finds a flaw in a product, and then you have to find everywhere that technology lives in your environment. PHP2Shell is a good example, with weaknesses in PHP itself and exposure that depends on version, configuration, and the specific flaw. You know about the company website. Do you know about the older portal, the vendor-managed app, or the PHP bundled inside some other product? Rapid7’s guidance warns that inventories can miss that kind of thing.
Meanwhile, the data behind vulnerability scanners is thinning. These are the tools that check your systems against lists of known flaws, and many of them pull their product details from the National Vulnerability Database (NVD), which NIST runs. Since April 15, NIST has prioritized NVD enrichment for known exploited vulnerabilities, federal software, and designated critical software instead of enriching every CVE. A scanner that depends on NVD’s product data alone may never connect a new CVE to what you run.
Fix: Fix is where the most time gets invested. Is there a supported patch? Can you test it? Will the application vendor support the change? Then someone has to approve it, and the affected system might handle billing or take customer orders, which tends to make everyone in the approval meeting very interested in the word “rollback.” Worrying about downtime is legitimate, because breaking a critical application hurts too. While that decision waits, though, the exposure stays open. And then there’s the question of whether the patch holds. N-able shipped four emergency hotfixes for N-central in five weeks this fall, after its first fix proved incomplete and new attack paths turned up in servers that were already patched.
This was a problem before AI sped anything up. Verizon’s 2026 DBIR found exploitation became the top way into breaches, at 31%. Only 26% of known exploited vulnerabilities were fully remediated, down from 38% the year before, and median time to resolution stretched from 32 days to 43. That data ends in October 2025, before most of this year’s surge, so remediation was already slowing before the flood arrived.
Verify: Verify is where “we sent the update” turns into “we confirmed we’re protected.” Did it reach every affected instance? Is the application loading the updated component? If a vendor manages it, what evidence did they give you?
Repeat: Repeat doesn’t take days off. You’re finding new exposures while old fixes wait for approval, and verifying yesterday’s deployment while today’s advisory lands. When work comes in faster than it goes out, the backlog grows, and even a strong team falls behind.
Where I Land
Right now, I think AI is making things more dangerous, and it’ll get worse before it gets better. Response still runs on knowing what you have, getting decisions made, deploying changes, and checking the result, and we’ve known those were weak spots for years. Faster discovery just raises the price of leaving them that way.
The best argument against me is that AI can help with response too, by digging through inventory, sorting findings, testing patches, and gathering evidence. Patch the Planet is a real attempt at that. If those gains start showing up consistently in deployed systems and shorter exposure windows, I’ll gladly change my mind. I just wouldn’t assume it happens on its own because the models get better.
So picture a critical vulnerability announced tomorrow in a technology sitting under several applications you’re responsible for. Could you answer these three questions with confidence?
-
Where are we affected? That covers the machines, the underlying technologies, and who maintains each one.
-
Who can authorize the fix? Settle emergency remediation authority before you need it. When access can go privileged in under three hours, the middle of an incident is a bad time to find out who signs off.
-
How will we verify it worked? Decide ahead of time what evidence closes the issue, including what you’ll accept from vendors.
Then ask your vendors the same three questions about their own products. Their answers will tell you more than any promise that they’re using AI for security.
This year’s jump in vulnerabilities shows how good AI has gotten at finding flaws. Fixing them still moves at the speed of approvals and maintenance windows, and that’s what I’ll be watching in 2027.