AI systems
Application Security Scanner
The expensive bug is the check that was never written.
It reads a web application and hunts for the page where a developer forgot to ask whether this user is allowed to see this, which is how most private data actually leaks.
Where it runsIt is not a website you visit. You point it at an application's code, it runs once as a job on a laptop or a server, and it hands back a report.
Who is this for?
Anyone shipping software that holds other people's data. The usual scanners look for bad code, and this gap is not bad code. It is code that was never written at all.
What the numbers mean
It gives the same answer every time, which is what separates a tool from a guess.
Everything it reported was real, so nobody burns a day checking a finding that was never a bug.
The checker is itself checked, because a checker that is wrong is worse than no checker.
How it actually works
Most security tools look for something bad in the code. This one looks for something missing, which is harder, because absence leaves nothing behind to search for.
The method is comparison. Work out mechanically how every page in the application protects itself, then line them all up side by side. Ten pages guard themselves the same way and one does not, and that one is the question worth asking.
Only the odd ones get sent to a language model, one at a time, with a single narrow question. The model is the slow and least repeatable part, so it gets the smallest possible job. Everything else is done by rules that give the same answer on every run.
That ordering is a measured result, not an opinion. Before it, the model read the whole application at once and surfaced one or two of the known bugs, a different one or two each time. After it, three separate runs produced the same four findings, and every one of them was real.
We test it against public websites that were built on purpose with known holes in them, so the right answers are published and anyone can repeat the run and check us. It is still in progress. It finds three of the four known holes, not four, and when we point it at sites built a different way it still comes back with nothing useful.
The leak that costs you is rarely a clever one. It is the single page where somebody forgot to check who is asking.
We build things like this. Tell us what you are stuck on.
What it is built with
- Python
- Docker
- Helm