A $20 AI Subscription Can Now Pentest Your SaaS Better Than You Can.

#ai #cybersecurity #saas #aitools

A French AI-coding YouTuber opened his Kimi K3 subscription, $20 a month, and went looking for security holes in another solo founder's app. He found some. He filmed the whole thing and posted it.

He had no written authorization and no pentest agreement, just a screen recording of someone taking apart a production app that wasn't his to touch.

The comments split into two camps within minutes. One side called it a straight up breach, unauthorized access to a system that isn't yours, the kind of thing that gets you a call from a lawyer. The other side called it a public service, free security testing for a community that mostly ships without ever checking for holes.

I'm not picking a side in that fight.

Skip The Ethics Fight. Look At The Price Tag Instead.

What he pulled off in that video would have cost, 3 years ago, years of accumulated offensive security skill. Today it costs a monthly subscription and a free afternoon.

That's the part worth sitting with. Not whether he should have posted it, whether he had permission, whether Medium or a lawyer somewhere agrees with his methods. The barrier that used to separate "random dev" from "person who can find your app's holes" has mostly dissolved. And it didn't dissolve just for him. I'm not naming him here either, no free advertising for a stunt like this, he doesn't need the traffic from me.

Any script kiddy with $50 and an afternoon can now point an AI at a SaaS and come back with something. Not a full breach every time. But enough to make "I couldn't afford to secure my app" sound less like an excuse and more like a choice.

The gate that used to require years of training now takes a credit card.

The Numbers Say This Is A Trend, Not A Stunt

This isn't 1 guy with a YouTube channel and good timing. It's measured, and the measurements keep landing in the same place.

A study reported by Hadrian, based on an academic paper published in February 2026, tracked a student with zero cybersecurity training who used CAI, an open source offensive security framework, and finished Austria's national CTF competition at the level of intermediate competitors who'd trained for months. No bootcamp involved, no mentor either, just the tooling doing the heavy lifting on reconnaissance. Think of it as XP farming, except the grind is automated and the character sheet fills itself in overnight.

Sonatype's data, cited by The Hacker News, shows the number of malicious packages published each year going from 55,000 in 2022 to 454,600 in 2025. Bugcrowd's 2026 research, cited by Penligent, puts AI adoption at 82% among bug bounty hunters already running it in their workflow, mostly for automation and code analysis. HackerOne's 2026 survey, cited by PentestPad, found that researchers using AI-assisted tools submit 28% more valid reports per month, with a higher severity distribution than manual work.

I've written before about how hackers already use AI better than you, and every one of these numbers just confirms it. Adoption isn't coming. It's already the default on the other side.

Here's the part that keeps this from sounding like a doomsday headline though. The same PentestPad piece points to an independent benchmark, CVE-Bench, where GPT-4 exploits 87% of known CVEs when given a description, but only 13% of real, undisclosed vulnerabilities on its own. The gap between "found something interesting" and "actually broke in" is still wide. AI is genuinely good at recon and discovery. It's still mediocre at carrying an exploit all the way through (more HAL 9000 quietly reading your files than Skynet taking over the mainframe). That's the honest state of things, not full autopilot hacking, more like a metal detector that occasionally finds gold.

Who's Liable When Audits Cost Nothing

This is where I think it gets interesting for anyone who ships a SaaS solo. The old defense, "I didn't have the budget or the skill to audit my own app," is losing ground fast. If a $20 subscription can surface real holes in an afternoon, "I couldn't afford it" stops being a technical limitation and starts looking like a decision not to spend an afternoon.

I think (and I could be wrong on where exactly this lands legally) that the standard of what counts as reasonable security effort just shifted for solo builders the same way the airbag shifted what counted as reasonable safety for a car maker. Nobody's saying you're legally required to hire a security firm before shipping a side project. But the argument that you had no realistic way to check gets weaker every month the tooling gets cheaper. I'm not a lawyer, this isn't legal advice, and I'd genuinely want to hear from someone who actually practices in this space before treating it as settled.

What I keep coming back to is a much older memory, completely unrelated to any of this. A long, long time ago, I used to disassemble games on my Amstrad, just for the thrill of cracking the copy protection. No purpose beyond the puzzle. And the annoying part was, the second a game was fully cracked, it lost basically all its appeal. The challenge was the point, not the game itself. I have no idea why that memory keeps surfacing while writing this one, but it does. Maybe it's because both involve the same kind of curiosity that doesn't care about permission, just about what's possible when you poke at something long enough.

The practical question for a solo founder isn't "am I legally liable," it's "do I want to find out the hard way, on camera, or find out first, on my own terms." Those are very different afternoons.

Audit Your App Before Someone Else Does

I already went deep on this in a separate piece, an audit of Supabase and RLS setups across a batch of vibe-coded apps. 91.5% of the apps checked in Q1 2026 had at least one exposed row-level security hole. Not exotic bugs. The default kind you get from shipping fast and never circling back, the classic "works on my machine, ships to prod, nobody checks" cycle every solo builder knows too well.

The method in my 3-prompt Supabase audit still holds. 3 checks, run this week, before anyone films themselves finding the gaps for you:

Every table with RLS enabled actually has policies attached, not just the toggle flipped on
No API route trusts a client-supplied user ID without checking it against the authenticated session
No environment variable with write access is exposed to the frontend bundle

None of this requires a security background. It requires 20 minutes and the willingness to actually look, the same willingness I talk about in Vibe Coding, For Real when it comes to the gap between a demo that works on your machine and an app that survives contact with strangers.

Sources

This post may contain affiliate links. If you click them, I might earn a small commission (costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure).