Agents that log in as two users to prove real data leaks
Viktor Bulanek is Founder & CTO of Penetrify.cloud, a security firm that runs AI agents to log into applications with real credentials and try to cross-customer boundaries. Their reports attach the exact request and the retrieved data the agent actually managed to get, which cuts through the hundreds of scanner ‘maybes’ and makes findings actionable.
Use AI to automate the testing work that benefits from scale and evidence: unsupervised agents that prove exploitability with real credentials make security testing both cheaper and more useful. At the same time, automate content and comms cautiously – Penetrify generated 46,000 pages across 24 languages and was demoted by Google until they deleted 98 percent and kept 588 human-reviewed pages.
What’s the quick origin story of your business, and what makes what you do genuinely different from the other options your customers are looking at?
I kept paying for penetration tests that were obsolete before I read them. A proper engagement costs somewhere between fifteen and fifty thousand dollars, takes weeks to schedule, and produces a PDF describing an application that has since been partly rewritten. Meanwhile the scanner we ran between those tests produced hundreds of findings and could not answer the one question that actually mattered, which is whether one customer could read another customer’s data.
So Penetrify runs AI agents that log into an application with real credentials, as two different users, and try to cross the boundary between them. The difference from most tools is what ends up in the report. A scanner tells you what looks wrong. Our agents only report what they actually managed to do, with the request and the retrieved data attached. That distinction sounds academic until you have spent an afternoon triaging four hundred maybes.
Where does most of your new business come from today, and where do you wish it came from?
Almost all of it comes from organic search, and specifically from unglamorous comparison pages. Things like "alternatives to X" or "what does Y actually cost". Nobody searches for a category they have not heard of, they search for the competitor they already know. Those pages took an afternoon each to write and they are the only thing that reliably brings buyers.
Where I wish it came from is recurring subscriptions rather than one-off purchases, and from integrations. Right now customers buy a scan, get value, and come back when they remember. I would rather be wired into the tools they already open every day so testing happens without anyone deciding to do it. That is a product problem more than a marketing problem, which took me embarrassingly long to admit.
What’s one thing you’ve actually handed over to AI in the last year, and one thing you tried it on and went back to doing the human way?
Handed over: the testing itself. Our agents run unsupervised for hours against live systems, and that genuinely works.
Went back: content. We generated about 46,000 guide pages across 24 languages, because we could. Google classified it as scaled content abuse and demoted the whole domain. We deleted roughly 98 percent of it and kept around 588 pages that a person actually reviewed. Traffic went up afterwards. The lesson was not that AI writes badly, it is that volume without judgement is a liability with a delay on it. Same thing with our media and sales writing. Anything meant to sound like me, I now type myself, because generated text is fluent and somehow says nothing that could only have come from one person.
Have you noticed a change in how customers find you or what they already know before they reach out? Is anyone arriving through AI assistants like ChatGPT yet?
Yes on the first part. People arrive having already compared us against two or three competitors, and they open with a specific objection rather than a general question. The research phase has moved off our website entirely, which means the job of our pages is now to be quotable rather than persuasive.
On AI assistants, I have to be honest: I have prepared for it more than I can prove it. We publish pricing as structured data generated from the same catalogue our billing uses, plus an agent skills manifest and an MCP server card, and our pages serve plain markdown to anything that asks for it. Whether that is producing traffic I genuinely cannot say yet, because AI assistants send visitors with no referrer. Anyone claiming precise numbers here is guessing.
If you were advising someone in your industry on all of this for the next 12 months, what would you tell them to actually do, and what would you tell them to ignore?
Do three things. Write the boring comparison content by hand, including the parts where a competitor is genuinely better, because that is what gets cited. Instrument your attribution before you spend anything on ads. We ran paid campaigns for weeks and later discovered our analytics were sitting behind a consent banner most people ignore, so every dashboard was measuring consent rates instead of customers. And pick one source of truth for your prices, because ours lived in four places and an AI will confidently quote whichever copy is stale.
Ignore anything framed as AI replacing a profession. In our field the useful change was narrower and much less exciting: work that used to be affordable once a year is now affordable monthly. That is a scheduling change, not a replacement, and the companies acting on it are quietly ahead of the ones debating it.
Thank you to Viktor Bulanek and the team at Penetrify.cloud for sharing what actually worked with Leaders Perception readers.
Want to share your playbook?
Leaders Perception publishes founder and operator interviews every week. A few questions, about five minutes, no calls. It is free and everyone we accept gets published.