How Voltti runs a security audit with AI, and where a human decides

A security audit with AI sounds like a button: code in, report out. That is not how we work. AI models and our own skills do the broad search through your code, your packages, and your configuration. Then a human reviews every finding and decides what is a real problem, how serious it is, and what you should do about it. Only then does it go into your report.

In this article, you will see how a security audit with AI works at Voltti when we review your app's code, from scoping to the recheck. In short, the five steps are scoping, a broad analysis with AI models and skills, a human review of every finding, a report with severity and fixes, and fixes with a recheck. You will also see why we give the search to models and keep the judgment with a human, and what you can prepare yourself.

A wide grid of dark pebbles under a band of blue light, with a few pebbles set apart on a light platform
The models search broadly. A human picks what matters.

Why AI does the search and a human makes the call

An AI-assisted code review finds suspicious spots in code quickly. Deciding which of them are real problems takes someone who understands the app and how people use it.

The makers of curl, a widely used open-source tool for network transfers, saw both sides of AI in security work within a single year. In May 2025, lead developer Daniel Stenberg wrote: “We still have not seen a single valid security report done with AI help.” People pasted code into a chatbot and forwarded the answer as a vulnerability, without checking whether it was true.

A few months later, the opposite happened. Security researcher Joshua Rogers sent lists of problems he had found with AI-assisted analysis tools. Together with a second researcher, he eventually reported over four hundred suspected issues. From the first list, the team merged about 50 bug fixes. The rest were partly false positives and partly minor issues the team did not think were worth fixing. Stenberg did not call the tools a revolution, but “an excellent new project assistant.”

The technology was the same. The difference was who used it. Stenberg later summed it up as “a clever person using a powerful tool.” The tool found the suspicious spots. Someone who understood the code decided what was real.

Research points the same way. When researchers had language models search for vulnerabilities in a strict test, even large models did not do much better than random guessing (Ding et al., 2024). Models are good at searching broadly and at connecting code across files. Whether a finding can really be exploited in your app is something a person has to check.

How does a security audit with AI work at Voltti?

An audit at Voltti has five steps. The models work mainly in step 2, the human in every step.

  1. 01

    Scoping

    We start with a conversation about your app. Which repository, which environment, and which parts do we review, and what is out of scope? If the app handles customer data, payments, or a language model, those parts get extra attention.

    You give read access to the code and the configuration. We do not need your production data. This is also where we agree on whether your code has to stay inside your own network.

  2. 02

    Broad analysis with AI models and skills

    AI models go through the code, the dependencies, and the configuration. They follow Voltti's own skills: fixed work instructions for each type of check. Think of secrets in the code, permission checks, or input that ends up in a database query. Or a language model that is allowed to do too much (more on that in prompt injection and API keys: the risks of AI features in your own app). That way, the check does not depend on how someone happens to phrase a question to a chatbot.

    We check against the OWASP Top 10 for web applications and, for apps with AI features, the OWASP Top 10 for LLM Applications. We also work with local models, so your code does not have to leave the network. Which model we use depends on the job.

  3. 03

    A human reviews every finding

    A person with 20+ years of experience in web development reviews every finding the models produce. Is it a real problem? Can a user or an outsider reach that spot in the code? How much damage could it do, and what is the best fix?

    We remove false positives. There are many: security firm SRLabs reported in June 2025 that around 80% of the findings from its AI tools were false positives. The OpenSSF guide on finding vulnerabilities with AI puts it plainly: a finding is only a potential vulnerability until someone checks it. What remains gets a severity. If we find a secret such as an API key or password, we report it right away instead of waiting for the report. You never get unchecked AI output.

  4. 04

    Report with severity and fix

    For each finding, you read where the problem is, what can go wrong, and how to fix it, with an example fix where possible. Each finding gets a severity: critical, high, medium, or low. That scale is similar to the CVSS rating scale, the standard that also scores known vulnerabilities.

    That way, you know where to start: with what someone could already exploit today. Your developer or your AI tool can start working with the report right away.

    A made-up example shows the format; it is not a finding from a client. On the order page, a logged-in user can open other customers' orders by changing the number in the web address. Severity: high. Fix: check on the server that the order belongs to the user who is logged in.

  5. 05

    Fixes and recheck

    You handle the fixes yourself, with your developer or your AI tool. If you want help, we include it in the quote. Then we check again whether the holes are closed.

    There is a reason for that recheck. An AI tool that fixes a bug sometimes only fixes the example from the report and leaves the same bug elsewhere. Or it creates a new bug somewhere else. That is why we review the changed code again after the fixes.

What the AI does and what the human decides

In short: the models search broadly, and the human decides. This table shows who does what for each task.

TaskAI models and skillsHuman

Scoping

No role

Decides with you what is in scope and which parts carry the most risk

Search

Go through the in-scope code, packages, and configuration using fixed skills

Looks closer at the parts with customer data, payments, or AI features

Suspicious spots

Flag anything that looks like a known bug pattern, including borderline cases

Checks whether the spot is really reachable and exploitable

False positives

Also flag issues that are not a problem

Removes them from the list

Severity

Do not know your app or your data

Decides the severity: critical, high, medium, or low

Fix

Suggest a fix

Picks the best fix and writes it into the report

Recheck

Go through the changed code again

Confirms whether the hole is closed

We use the same principle for our own work. Every content change that our own Hermes installation at Voltti makes starts as a ticket and ends with a verification report in that ticket. People check the finished work and adjust where needed.

Why we also work with local models

We also work with local models because an audit means reading a company's source code. And not every company wants to send that code to an outside AI service. That code often contains the database structure, internal names, and sometimes a key by accident. With local models, your code does not have to leave the network. We agree on the right setup during scoping. You can read how this works in general on our page about local AI for businesses.

Local does not mean the judgment goes away. A human still reviews every finding.

What you prepare yourself

You do not need to do much before an audit starts. This does help:

  • Read access to the repository, including its history. Old commits sometimes still contain a key.

  • A short explanation of what the app does, who uses it, and which roles exist.

  • The services the app talks to, such as a database, a payment provider, a language model, or email.

  • Which parts are new or were built by an AI tool, if you know.

  • Who will make the fixes: your own developer, the agency that built the app, or Voltti.

What an audit is not

A security audit looks at code and configuration. A pentest, where someone attacks your running app from the outside, is something else and is not included. The two complement each other. We do not issue a certificate or seal, and no audit can promise that an app is 100% secure. What you do know afterward is which holes are in your app right now, how serious they are, and how to close them.

Want to look for yourself first? Start with the vibe coding security checklist and the 7 most common security flaws in AI-generated code. You can read why an audit is needed in why an app built with AI needs a security audit. Want your app reviewed? Take a look at the security audit by Voltti.

Portrait of Seppe Gadeyne

Seppe Gadeyne is a developer and the owner of Voltti. He sets up AI agents for companies in Belgium and the Netherlands.

Updated on

Frequently asked questions

01How does a security audit of an app work?

It has five steps. First scoping, then a broad analysis with AI models and our own skills, and a human review of every finding. After that come a report with severity and fixes, and a recheck after the fixes. The models do the search. A human decides what goes into the report.

02Why don't you let the AI do everything?

A model also flags issues that are not a problem. It also misses bugs that only show up when you know how people use the app. A human checks whether a finding can really be exploited and how serious it is. That way, you do not get a list full of false positives.

03Does my code have to go to an outside AI service?

No, it does not. An audit means reading your source code, which often contains the database structure and internal names. That is why we also work with local models, so your code does not have to leave the network. We agree on the right setup during scoping. We do not need your production data.

04What do you do if you find an API key or password?

We report it right away, without waiting for the report, and we advise you to replace the key with the provider. Just removing the line from the code is not enough, because the key is still in the history of your repository. During the audit, we also look through that history, since old commits sometimes still contain a key.

05What does a security audit cost, and do you help with the fixes?

You always get a custom quote. The price depends on the size of the code, the number of integrations and environments, and whether you want help with the fixes. If you want that help, we take on the fixes and include them in the quote. A recheck follows the fixes. We work with businesses in Belgium and the Netherlands.

Which app do you want reviewed?

Discuss your pilot