TL;DR
- The real vulnpocalypse is the AI-built software nobody will maintain. Old kernel bugs have maintainers and patch cycles. The apps your teams are building with AI right now don’t (or won’t in the future) and they’re repeating the same old coding mistakes.
- Don’t block it. Build the paved road. Security teams that make the safe path the easy one will help the business move faster and keep the sprawl in check.
- Take it to the board as a basics problem. Vulnerability management and supply chain oversight need more investment than they get today, starting with knowing what’s been built, who owns it, and what it can reach.
The vulnpocalypse has officially been coined, and I don’t love the phrase. The problem is real, but I think most of the conversation is pointed at the wrong software, and the term itself isn’t that helpful.
In my last post, I mentioned the 24 Linux kernel zero-days found during the [un]prompted CTF, which is exactly the kind of headline that gets people worried about AI digging decades-old bugs out of foundational code. Still, that code has maintainers, release processes, and a long history of getting patched, so it’s the more fixable problem.
What worries me more is the software being written right now by people who’ve never shipped code before, and that nobody is going to maintain once it’s running.
SaaS was a reaction to this exact problem
A former colleague and I have a long-running joke that we’re building tomorrow’s legacy software today, and it describes just about every company now. And more importantly, every function within a company is now contributing to that problem.
Businesses have built their own apps on low-code tools since the early ’90s, and plenty ended up running critical processes on homegrown Lotus Notes, Power Builder, or AS/400s for decades. Once your workflows depend on something like that, getting out takes about as long as getting in did.
SaaS was the industry’s answer: rent the HR system or the task tracker, and hand off the maintenance along with it. The tools were never quite what you needed, and your users absorbed the gap, but that still beat maintaining code you didn’t have the people to maintain.
AI has removed the reason that trade made sense. Anyone with domain expertise can now build exactly the tool they want—a COO can build a customer success system instead of buying one, and SOC analysts are building agents that run on their own laptops. What used to be a major engineering project, you could knock out on a Tuesday.
That’s why forward-thinking companies are moving budget from software licenses to AI tokens, and for the most part I think it’s the right call. The trouble is what it leaves behind.
Who owns the app nobody remembers building?
Fast-forward two years. Someone on your marketing team built a great app, the whole team runs on it, and then they take a new job. The model underneath has changed, the framework is three versions behind, and no one is responsible for keeping it current. In most companies, it either sits there aging or lands on a stretched IT team with no tooling built for the job.
Multiply that across every team in the business. You’ll cut seat-based licenses, and in their place you’ll get a sprawl of apps nobody owns. Calling that shadow IT undersells it. It’s new IT: freshly built, running real workflows, and mostly invisible to the people responsible for securing it.
There’s a second problem layered on top. LLMs learn from existing code, so the vulnerabilities and bad habits already out there get repeated, starting with credentials left exposed where they shouldn’t be. That particular mistake is already showing up in our data. In Q2, exposed cloud secrets accounted for 29% of the cloud infrastructure incidents the Expel SOC handled, compared with 3.2% for server-side vulnerabilities. AI-generated code didn’t create that problem, but it’s very good at repeating it. Most people building with AI don’t read the code it writes, because what they care about is whether the thing works.
Put those together and we’re creating a bigger problem than the one we’re solving. The old bugs in old code may turn out to be the easier fix.
Build the paved road
So what does a security leader do when every team is building its own apps? Trying to stop it is a losing fight, and the business is making the smarter play anyway.
If you have a product security team, point it here and start building paved roads (the classic SRE move of making the safe path the easy one). In practice, that might be an agentic harness that runs security checks on whatever people are building, plus continuous rebuilds of their software, the same way you already handle the rest of your stack.
That work isn’t on your roadmap, and nobody is eager to pick it up, but the mold is already growing in the corner. The longer you wait to look at it, the bigger and hairier the cleanup gets.
Be clear about scope, too. Security can’t maintain every app people build, but it can keep those apps from turning into infrastructure and security problems. Be upfront that the road has a cost, and that anyone putting an app where others depend on it should pay a small tax to use it. Done well, this gives the CISO a chance to help the business move faster, and “here’s how to ship it safely” lands a lot better than “no.”
Know what that cost looks like. Some apps cost nothing to run once they’re built, while one with built-in summarization or an agent enriching alerts around the clock burns tokens every day, and that operating expense compounds. You have more control over it than most people assume. Plenty of everyday work doesn’t need a frontier model, and a lot of what I used to send to the most expensive tier runs fine on a much cheaper one now.
The long-term answer is likely software factories that continuously rebuild and update codebases in production. We’re still in the early days of loop- and graph-driven development, though. In most organizations, compliance, release management, and operations have a lot of catching up to do before that model works.
Take it to the board
The best CISOs are already raising this with their boards. Shadow IT has always needed managing. What’s changed is that most companies, in the rush to move faster with AI, have relaxed the guardrails that used to keep it in check. We’ve made that trade at Expel too, adopting MCP servers early because the speed was worth it to the business, and we made that call with our eyes open. Moving fast is a reasonable choice, as long as someone is accounting for the risk.
Vulnerability management, both cleaning up what already exists and governing what gets built next, needs more investment than it has today. That includes the software supply chain and where your data, workflows, and identities come from. The plain version tends to land with any board: a lot of people who’ve never written software are now shipping code in our environment, and most of them don’t think of what they’re doing as shipping code.
Ask better questions about what you’re building
You don’t need a finished program to get started. You need honest answers to a few questions that most teams can’t answer yet:
- How many apps are running in our environment that nobody in IT or security knows about?
- Who owns each one if the person who built it leaves tomorrow?
- What data, credentials, and systems can those apps reach?
- If we were starting today, would we let this get built the same way?
This is as much a context problem as a code problem, because you can’t make good security decisions about software you don’t know exists. Every company is carrying this risk right now. The choice is whether you decide how much of it to carry, or find out in two years what the default was.
