Most applications never get tested. The majority of web applications have never had a penetration test. Not because teams don’t care, but because the economics don’t work.
A manual web app pentest costs $10,000 to $20,000, takes weeks to schedule, and depends on a small pool of senior testers who are booked out months ahead. So testing turns into a once-a-year event for one or two critical assets, while code ships every week.
That breaks in two ways. First, coverage: most of the attack surface never gets looked at. Second, staleness: a point-in-time report is stale the day the next release goes out, and most incidents come from changes made after the last test.
What AI actually fixes is throughput, cost, and frequency. If a real test can be kicked off in minutes and delivered in days at a fraction of the traditional price, you can test everything, and test it often. The capability question I’ll get into later, but the economics question is already settled: thorough testing being a scarce, once-a-year luxury is the part that was actually broken, and that part is fixed.
Fair question, because a lot of “AI pentest” products are exactly that: a scanner with a new label. The difference isn’t speed. It’s whether the system reasons and validates.
A DAST scanner (or vulnerability scanner) matches signatures and known patterns and hands you a list of potential issues. It doesn’t understand the application, doesn’t chain steps together, and doesn’t confirm whether anything is actually exploitable.
That’s why scanner output is full of false positives. A penetration test works the other way. It explores the application like an attacker would, chains weaknesses together, attempts real exploitation, and confirms each finding is reachable.
We built Vana to work like a tester, not a signature engine. It authenticates into the application, works out how to abuse it, attempts the exploit, and validates before anything goes in the report.
What you get is a set of confirmed, reproduced findings, the same artifact a human tester delivers. Not a scan dump.
So the test is simple: does it validate and chain, or does it just enumerate faster? If it only enumerates faster, it’s a scanner with better branding.
Our company is built on proving AI can test as effectively as humans.
Technically, every single operation in testing can be done by AI today, if it’s been given the correct tools to do it. Reconnaissance, authentication, exploitation, chaining, validation, reporting. There’s no step in the methodology that’s inherently human.
Where AI can lag is speed of adaptation. When a new vulnerability class or exploit technique drops, an experienced person reads the write-up and is using it against targets within minutes. An AI system needs its training and tooling loop to catch up first, and that loop, however fast it’s getting, is still slower than a sharp person reading a proof of concept over coffee. So for anything brand new, the person has a window of advantage.
But this isn’t always the case even now, since there are already tools used to counter this exact problem. RAG or Retrieval-Augmented Generation is a core part of solving this, allowing an AI to research and absorb live information the second it starts it’s testing. So, It’s an engineering problem, and the loop keeps getting shorter. For the overwhelming majority of what a real engagement covers, the vulnerability classes that actually get organizations breached, the gap in capability is gone. The only gap left is how fast we adapt to brand-new techniques, and that’s closing.
Four ways:
Validation. DAST flags potential issues and leaves you to sort real from noise. Vana attempts to exploit what it finds and only reports what it can confirm. That cuts false positives and gives you findings you can act on instead of a triage backlog.
Chaining. A scanner tests issues in isolation. Vana reasons about attack paths and chains vulnerabilities the way a real attacker does. That’s where the high-severity findings come from, and it’s exactly what scanners miss.
API coverage. Vana tests the web application and its APIs. APIs carry a huge share of modern risk and they’re where scanners are weakest, because meaningful API testing needs authenticated, stateful, logic-aware interaction, not pattern matching.
The operational model. It’s self-serve with flat, predictable pricing. Drop in a URL, get an enterprise-grade report in about four days. That means you can test every asset and every material release, not one asset once a year. And the deliverable is a real Pentest report with reproduction steps and remediation guidance, suitable for compliance and diligence. Not a report full of maybes.
The routine, repeatable parts. Testing common, well-documented web and API vulnerability classes. Running standard tooling through a standard methodology. Manual recon and enumeration. Writing up standard findings by hand. That work is being automated now, and the pace is increasing. Being a scanner operator is not a durable career.
What grows in value is everything above that: adversarial creativity, threat modeling, business-logic analysis, exploit development for novel classes, deep cloud and identity and API architecture expertise. There’s also a new attack surface that barely existed a few years ago: AI systems themselves. Prompt injection, model and agent abuse, securing agentic workflows. And there’s a new discipline in directing and QA-ing AI-driven testing, the senior human who orchestrates the machines and validates their output.
So the profession isn’t disappearing. It’s moving up the value chain. The floor rises, and the ceiling matters more than ever.
Both, with the weight shifting to continuous. The annual-pentest-as-checkbox model doesn’t match modern release velocity, and it doesn’t match where regulation is going either. NIS2, and DORA in financial services, treat risk management as an ongoing obligation, not an annual event. If you test in January and ship two hundred times before the next assessment, that report says very little about your risk in October.
The practical setup is layered. Keep periodic, deeper engagements, including human-led work, where they add real value or where they’re mandated. Underneath that, run a continuous automated validation layer as the always-on baseline: test on every material change, whether that’s a new feature, a new endpoint, or the run-up to an audit or investor diligence. Continuous validation used to be economically impossible, which is the only reason “periodic” became the standard. AI changes the economics.
The point is to satisfy the spirit of continuous risk management, actually knowing your posture at any given moment, not just the letter of an annual requirement.
Two stand out.
The first was a chain. Vana found a LaTeX injection on an endpoint and chained it through a Tomcat auto-deploy to modify a PDF. Processing that PDF executed code that changed where a file got written, and that gave Vana a reverse shell. Full remote code execution, built entirely out of pieces that looked harmless on their own. That target had already been tested by more than fifty senior pentesters and a range of tools before us. None of them found it, because the vulnerability wasn’t in any single component. It was in how they combined. That’s exactly what signature-based tools and time-boxed manual tests walk past.
The second was on a production site where every prior test had reported no high-severity issues. Vana found a hidden XSS that had been sitting there through all of them, and used it to take over the site’s chatbot completely. A clean high, on an application everyone else had signed off as clean.
Both make the same point. The findings that matter most are usually the ones that require chaining steps together, or reaching parts of an application that shallow, point-in-time testing never touches. That’s where reasoning-driven testing earns its place.
Stop equating “we passed our annual pentest” with “we’re secure.”
Different statements, and the gap between them is where breaches happen.
Stop confusing a vulnerability scan with a penetration test. A scan is a list of possible issues. A Pentest is validated, exploited, prioritized findings.
Buying the first and believing you got the second is one of the most common mistakes I see, and one of the most dangerous.
Stop leaving assets untested because testing is “too slow or too expensive.” That was true for a long time. It isn’t anymore, so it’s no longer a defensible reason to fly blind on most of your attack surface.
Also, stop testing one asset once a year while every other application, API, and release goes unchecked. The fix is simple to state: go from periodic to continuous, validate instead of enumerate, and test what ships.
Put your brand and expertise in the spotlight with one of our carefully crafted sponsorship packages. Whether it be a speaking role, a delegate package for your team, logo exposure, or the opportunity to bring your current and potential clients along to the event, we have got you covered with something that will genuinely help you get deals done at our events.
Join us in uniting for a safer tomorrow!
