## 0006 / 2026-08-31

Cybersecurity, in the Agent era

$1,000 spent on AI inference is most likely enough for a bad actor to compromise any Software you (or your team) are currently building.

They can use unrestricted models. They can use Chinese models. They can use whatever. And if you haven't spent a proportionate amount of tokens on your own side, you're probably going to get hacked in some way.

Maybe not full production database access. Maybe they find something. Maybe they DDoS you. Addresses leaked. Payments. API keys. Reputational damage. I don't fucking know. Depends how far they get.

All it takes is one bad actor with a couple hundred dollars.

And this is not some future sci-fi thing. We are looking at the first generation of truly powerful unrestricted models that run on a mid-level computer. A Qwen 3.8 27B runs on 24 gigs of RAM, maybe 32. That's enough to do a lot of damage to 95%+ of software that does not have good security.

Now imagine a few months from now. A few years. We started with unrestricted small models. Now we're getting unrestricted medium-sized models. GLM 5.3 Flash is already one of the default models on cybersec benches, even compared to frontier. That's the restricted version. Unrestricted DeepSeek, unrestricted GLM... one person with a normal budget can completely fuck you up.

It's only the number of tokens you throw at it. $100. $1,000. $10,000. That's the main variable.

This is one or two generations of models away from being everybody's problem. Every small site. Every piece of software that isn't spending money to attack itself.

So if you haven't tried to attack yourself, you're not going to be able to defend yourself.

This is the proof of work in the age of AI.

Bitcoin works because the energy cannot be faked on the protocol. Same principle. How much have you spent trying to break your own software? That number cannot be faked either.

The practice is simple. You run unrestricted models against your own app. You set a budget. You only give them what an attacker would know looking at the site -- no internal info, no source, no secrets. Then you fix what they find.

That's it.

People think the first targets are Google, Facebook or Apple. But that's not the case... Those companies are fine.

They get early access to new AI models. They work with OpenAI and Anthropic directly. They have tens of thousands of engineers. When Mythos got announced, they handed it to these companies, and most of them still found security issues in their shit.

The problem is everybody else. Small and medium-sized businesses. Anyone building software.

I know this because we've already felt a weak version of it. Last year, when we put Vectal on Product Hunt, we immediately got like 15 emails about "vulnerabilities." Most of them were fake. One of them had something worth patching. He didn't have real access, but it was enough. Launch something public and people start testing your stuff the same day.

It's a matter of time. Could happen tomorrow. Could happen next month. But it will not take a year. This is right around the corner.

And if you skip this, you don't save the money. You just pay later, in time. Someone gets in, and then you're spending all day, every day trying to get them out, recover your access, and clean up a complete headache. Same as skipping accounting. Same as skipping legal.

This is going to be one of the main costs of building any software. How many tokens have you spent on self-pentesting. That's cybersecurity in the agent era.

All articles