Blog / AIpento

SAST, DAST, pentest, AI pentest: what each one catches

Four names for testing an application, and they find different things. Attackers now run AI against your app continuously, so the question is no longer which test is most thorough but which one can keep pace. A plain guide to what each does, what it misses, and where AIpento fits.

If you are buying application security for the first time, the names blur together. Here is what each one actually does.

SAST: reading the code

Static application security testing reads your source code without running it. It is fast, it can run on every commit, and it points at the exact file and line. It is good at patterns: a key pasted into a file, a query built from a string, user input written into a page as HTML.

Its limit is that it cannot see how the running app behaves. It does not know that a route is protected by a proxy in front of it, or that a check in another service makes a line safe. So it raises some things that turn out to be fine, and misses problems that only exist when the pieces run together.

DAST: testing the running app from outside

Dynamic application security testing sends requests to a running application and studies the responses, the way an outsider would. It finds what only exists at runtime: missing security headers, a misconfigured server, an endpoint that answers when it should not.

Its limit is depth. A scanner works through a list of known tests. It does not understand that a "share" button and an "export" feature can be combined, or that the second user's invoices should not be visible to the first.

Penetration testing: a person attacking on purpose

A penetration test is an authorised attack by skilled people. They map the app, form a plan, try what a real attacker would try, and chain small weaknesses into a serious one. It is the gold standard for depth, and auditors for SOC 2 and ISO 27001 often ask for one.

Its limits are time and cost, and the AI era has made both worse. A test takes weeks to book, days to run, and describes the app as it was on those days; by the next release parts of it are already out of date. The harder change is on the other side: the attacker is no longer working at human speed either. The same kind of model that writes your features is being pointed at them, mapping a surface and chaining small weaknesses into a breach in minutes, continuously, at no real cost. A quarterly test was a fair answer when attacks were handmade. Against an opponent that never sleeps and never books a slot, it is a snapshot you are hoping stays true.

AI penetration testing: the method of a pentest, at the pace of software

An AI penetration test uses AI agents to do the work a tester does: map the attack surface, plan, test, chain and report, against an app you own and have authorised. Done well, it brings the depth of a pentest to every release instead of once a year.

Done badly, it produces confident nonsense. The question to ask any AI pentest vendor is simple: what happens to a finding the system cannot prove? In AIpento it is withdrawn, and every finding you see carries the request and response that proved it. We explain how in Reproduce or withdraw.

Where AIpento fits

AIpento combines the first and the last. The readiness scan reads your code (SAST-style). The AI penetration test runs against a staging copy of the app you authorise, and keeps only what it can prove. Then AutoFix writes the patch and accepts it only when your test suite passes three times, each from a clean checkout.

The scan also covers what a classic SAST tool leaves to other products: more than 200 kinds of leaked secret, npm and PyPI packages with known vulnerabilities (checked against the public OSV database), and infrastructure files such as Dockerfiles, CI workflows, Terraform and Kubernetes.

What it does not do: it does not check packages from other ecosystems yet, and it only tests targets you have authorised in writing, never a production system. We say so on our security page, because a tool that is vague about its limits is hard to trust about anything else.

Keep reading