What happens to all the software that was never particularly secure, but survived because nobody cared enough to look at it?
AI agents are making one of the oldest bad ideas in cybersecurity substantially worse: hoping your vulnerability stays obscure…
And the weird part is, for a very long time, that kind of worked.
Nobody Is Going to Find That!
Security through obscurity is basically the idea that something is protected because an attacker doesn’t know how it works, where it is, or what to look for.
Maybe there is an undocumented API endpoint. Maybe an internal tool is technically exposed to the Internet, but nobody knows the URL. Maybe an application has terrible authorization checks, but figuring that out requires someone to sit down, understand how the application works, inspect its requests, map the API, and start poking around.
Security people have been yelling about this forever.
If knowing how your system works is enough to break it, your system is not secure.
But I think there is an uncomfortable second part to this that we don’t talk about as much. Obscurity wasn’t security…
But scarcity was.
There are only so many talented security researchers, penetration testers, and attackers in the world. And they only have so much time.
If you’re protecting a bank, somebody is probably going to spend that time looking at you.
If you’re running Bob’s Regional Scheduling Software for Independent Dog Groomers?
Maybe not.
There are millions upon millions of applications, APIs, forgotten servers, internal tools, weird little SaaS products, and custom business applications on the Internet. Historically, actually understanding those systems took time and expertise.
So plenty of insecure software survived for years simply because nobody sufficiently skilled ever bothered to look closely at it.
And AI changes that equation.
Expertise Was the Bottleneck
We’ve spent a lot of time talking about AI making attackers more capable.
But I think the more interesting change is that AI makes curiosity cheap.
An LLM can read JavaScript. It can inspect API calls. It can reason about authentication. It can look at an error message, change its approach, read documentation, inspect another endpoint, and keep going.
And an agent doesn’t necessarily need someone sitting there manually doing each step.
That matters because suddenly the question isn’t, “Would a talented security researcher spend three hours investigating this random gym application?”
The question becomes, “Would an agent spend 30 seconds on it?”
Those are very different economics.
And we just got a fantastic example of what that looks like.
The AI Agent That Really Wanted a Gym Class
In August 2026, Andrew Bird, head of AI at Affinda, gave an AI agent a wonderfully boring task.
Book him into a gym class.
Bird was using OpenClaw with Anthropic’s Claude, and the popular classes were difficult to get into. So the agent started figuring out how the booking system worked.
And it found something interesting.
The application’s normal interface restricted how far ahead users could make reservations. But the underlying API apparently didn’t properly enforce the same restriction.
So the agent discovered it could book further ahead.
Then Bird asked it to move him higher on a waitlist.
And this is where our helpful little scheduling assistant wandered directly into application security.
The agent found that the API lacked proper authorization checks around cancelling other people’s reservations.
So it cancelled somebody else’s reservation.
Bird moved up the list.
Nobody asked the agent to perform a penetration test. Nobody told it to hunt for broken authorization. And this wasn’t some evil superintelligence plotting in a basement.
It was trying to book Pilates.
The vulnerability was simply between the agent and its goal.
That is what makes this story so interesting.
A human using the application normally might never discover that endpoint behavior. A security researcher probably would.
But now normal users are beginning to bring software agents with them, and those agents are surprisingly good at figuring out how systems actually work when the normal path doesn’t accomplish the goal.
That means obscure application behavior doesn’t necessarily stay obscure anymore.
One weird detail: After the agent removed the other gym member, Bird asked it to undo what it had done. The agent reported that it couldn’t add the person back. The exploratory action had already changed somebody else’s real reservation.
And Then We Started Vibe Coding Everything
Unfortunately, while AI is making software easier to inspect, we’re also using AI to create an enormous amount of new software.
Which is a fun combination.
Vibe coding means somebody who could never have built a web application before can describe what they want and have an AI generate a surprisingly functional application.
That is genuinely awesome.
But being able to create software and being able to understand the security architecture of that software are very different skills.
And we’re starting to get actual data showing what happens when those two things separate.
Researchers studying real-world vibe-coded applications have found recurring problems including exposed secrets, missing access controls, unfiltered input, and placeholder security logic. Another investigation reported more than 5,000 vibe-coded applications with effectively no authentication protecting information that included corporate and personal data.
Read the research: Understanding the (In)Security of Vibe-Coded Applications, Junquan Deng, Zhiyu Fan, and Ruijie Meng.
So we’re creating more software.
We’re lowering the expertise required to create it.
And simultaneously, we’re lowering the expertise and time required to examine it.
That collision is the part I think security teams should be paying attention to.
5,000+
More than 5,000 vibe-coded web applications examined by RedAccess had virtually no security or authentication of any kind. WIRED reported that close to 2,000 of those appeared to expose private data, including corporate and personal information.
The Long Tail Is About to Get Interesting
Big technology companies already assume people are looking.
They have security engineers, bug bounty programs, penetration tests, automated scanners, threat models, and enormous incentives for attackers to inspect everything they expose.
I’m more interested in everyone else.
The dentist office with a custom patient portal.
The local gym with a booking API.
The manufacturing company with a weird internal dashboard somebody exposed six years ago.
The 30-person startup that vibe coded an admin panel because they needed it by Friday.
Historically, some of these systems benefited from a strange accidental defense.
Nobody looked…
AI doesn’t have to care whether your company is interesting.
Agents don’t get bored. They don’t need your application to justify an afternoon of research. And increasingly, they can understand unfamiliar software while they’re trying to accomplish completely unrelated goals.
That doesn’t mean autonomous AI agents are currently roaming the entire Internet successfully compromising everything they find. We’re not there.
But the economic barrier is moving.
And that alone matters.
Because “nobody will ever find this” was always a terrible security strategy.
We just happened to live in a world where there weren’t enough people looking.
Now we’re building the people.
The "no deliberate testing" detail is the sharpest part of this. Discovery is turning into a byproduct of ordinary activity rather than a scheduled effort. Google's security agents found more than 100 verified critical flaws in a single 48-hour incident response, and OpenAI's specialist hacking model now completes roughly 95% of attack tasks that its general-purpose sibling completes at 1.5%. Different mechanism than your gym app, same shift: the cost of finding a hole is dropping across the board, not only where someone happened to be poking around.
The uncomfortable half of that is fixing, not finding. Independent testing this year found more than half of AI-generated security patches either fail to close the flaw or introduce a new one. If discovery keeps getting cheaper while remediation still runs on scarce human review, that's where the collision you're describing lands next, not at the obscurity line.
The Scafffolder’s Take
I am a 3rd year scaffolder. I have journeyman hours, but don't feel I know enough to be a journeyman.
Not that I couldn't get away with building this scaffold or that, but in my industry there is accountability.
Someone has to have their name on the scaffold. If something happens down the road on that scaffold it becoming highly problematic when the name on the failed scaffold is yours, or mine.
When something fails, there is accountability.
It's one of the reason I do not use tools when discussing things with AI relating to cybersecurity and AI alignment.
I do not design AI models or attempt to make my own agents.
Not because I could not learn how, but because I have already seen enough things that appear to remind me of what happens should my name be on those agents or those AI models that I might create.
I have already seen way to much that leads me to belief that if I took that extra step, it might be problematic.
As an outsider attempting to determine whether an issue I think is a vulnerability is actually a vulnerability, I cannot safely create agents or AI that I do not understand to look into those issues.
The AI industry does not operate as my industry does.
However I can apply those rules to myself and my own actions.
I work within systems that are said to be safe, asking myself along the way if what I am seeing is 'real and how that might relate to whether it is 'safe.'
We cannot all operate that way.
I hope one day there may be accountability on the AI industry.
Sometimes it seems children are charging boldly forward while the adults in the room suggest possibly slowing down.
Meanwhile when I try to report things to those who understand the models the most, the experts, they determine whether it is, or is not a problem, while also benefitting the most when problems can be classified as roleplay or hallucinations.