Opinion
Can penetration testing be 90% cheaper?
A penetration test quickly costs six thousand euros. For a complex environment that runs to well over a hundred thousand. Even a modest test of two to three thousand euros is a serious expense for the average SME, certainly if it has to be repeated every year. For a long time that was a choice: you did it, or you did not. Unless sector regulation required it, in which case there simply was no choice. With the arrival of the European NIS2 directive, implemented in the Netherlands as the Cyberbeveiligingswet, that is changing. Thousands of organisations now fall under legislation that requires technical security measures, including testing their own resilience regularly. Many organisations translate that into a penetration test.
The pentest market does not scale with that demand. Pentesters are scarce, training takes years and hourly rates are rising. You can wait for the market to adjust, or you can try to solve the problem differently. We chose the latter.
Guardian360 is one of thirteen European organisations that received a grant under the European Union’s CYSSDE programme to make penetration testing more accessible. Our project is called AI-SCALE. Our ambition: to offer professional penetration tests at around ten percent of the usual price, by letting an AI agent do most of the work, with human validation of the findings. This article is not about how well that is going, because we are not there yet. It is about what we are running into, what we have learned and which questions we cannot yet answer.
What we are building
The technology is built in three layers. The first layer is continuous discovery: our existing Lighthouse platform continuously maps the attack surface, both external and internal. The second layer is the AI-driven assessment: our own AI Cyber Security Agent that generates and runs pentest scenarios, tuned to the sector and the technology stack of the organisation. The third layer is the human review: a senior certified pentester from one of our pentest partners validates the findings before the report is delivered.
That three-layer structure is not a compromise but a deliberate choice. The AI does the work that scales: scanning broadly, working through scenarios, documenting findings. The human does the work that requires judgement: deciding whether a finding is a real risk, assessing how vulnerabilities combine, and deciding whether exploitation is responsible.
When the project concludes in March 2028, the AI agent, the per-sector scenario libraries and the full methodology will be published as open source. That is a commitment in our project application, and it is also our conviction: professional security tooling should be testable, not a black box.
What is a penetration test, really?
Before we can talk about a 90% cost reduction, we have to agree on what we are trying to make cheaper. And that is where the first problem begins. The term “penetration test” has no single definition. For one supplier it is an automated vulnerability scan with a report. For another it is weeks of manual work by an OSCP-certified specialist who actually tries to break in.
Compliance frameworks only help so far. NIS2 asks for “testing the effectiveness of security measures” but does not prescribe how. ISO 27001 refers to technical vulnerability assessments without specifying the depth. DORA is more precise about threat-led penetration testing (TLPT), but that only applies to the financial sector.
In the Netherlands there is a concrete quality standard: the CCV Keurmerk Pentesten, a certification mark for penetration testing. The Centre for Crime Prevention and Safety (CCV) developed a certification scheme that sets out what a pentest is and what it is not. The definition is unambiguous: a pentest is “a manual assessment that attempts to penetrate a system as deeply as possible”. Vulnerability scans, however advanced, are explicitly excluded from the mark. Tooling may be used as an aid but is, in the CCV’s words, “not leading in the execution of a pentest”. All findings from automated tests must be manually verified by a qualified pentester.
On top of that comes a hard quantitative requirement: the qualified pentester must deliver at least fifty percent of the total test time. That pentester must also hold a recognised practical certificate, have at least one year of work experience as an IT professional (not as an intern), and a certificate of good conduct (VOG) no older than three years.
For anyone who wants to offer AI pentesting, this mark raises the central question: if the AI agent does most of the work and the human validates, is that still a pentest in the sense the CCV means? The honest conclusion is that the three-layer model fits the spirit of the mark conceptually: tooling as an aid, the human expert ultimately accountable. But the fifty-percent rule and the definition as a “manual assessment” sit uneasily with the ambition of a ninety percent cost reduction. If the human pentester has to deliver half the hours, then by definition you do not reach ten percent of the usual price.
In the AI-SCALE project we therefore use a broader definition: a pentest is a structured assessment that goes further than running an automated scan. It includes scenarios tuned to the sector and the tech stack, covering misconfigurations, authorisation errors and business-logic weaknesses, and it ends with human validation.
The limit of automation: exploitation
One of the first conversations we had with our pentest partner was about exploitation. May the AI agent actually exploit a vulnerability it finds, to prove the risk is real? The answer was crystal clear: no, in any case not fully automated.
That sounds like a limitation, and it is one. But it is a deliberate limitation. An AI agent that exploits vulnerabilities in a production environment without human oversight is a risk in itself. In our model the agent operates mostly in read mode: it explores, identifies and reports. Every action that changes a system requires human approval. There is a kill switch, there are Rules of Engagement per assignment, and there is extensive audit logging.
For a CISO that may be more reassuring than a fully autonomous agent. But it also means the cost reduction will be limited. As soon as a human pentester has to be significantly involved in the validation process, a saving is still possible, but how large it will be is still unclear.
False positives cost validation time
The AI agent is strong on breadth. Misconfigurations, outdated components, default settings and known vulnerability patterns it finds consistently, including in the places where a human looks less sharply after a few hours. But breadth has a downside: the agent produces more findings, and not all of them are relevant.
If the AI produces more noise than a human tester, you shift the hours from the test to the validation. The human pentester then spends more time assessing and rejecting findings than discovering them. Recent benchmarks confirm that false-positive rates are the big bottleneck for automated pentest tools. Time gained at the start of the test you lose again at the end.
Business logic and chained attacks
The agent finds individual points of attention, but does not always see that together they form a serious attack chain. An authorisation error between roles, a payment flow that skips a step under certain conditions, an API that is safe in isolation but enables a privilege escalation in combination with another endpoint: these are exactly the findings organisations hire a pentester for. And they are exactly the findings AI cannot do on its own for now.
That is not a matter of better models. It requires understanding what an application is supposed to do, in the context of the business process it supports. That is domain knowledge, not pattern recognition.
The assumptions behind the cost model
The target of ten percent of the usual price, roughly six hundred euros per test instead of six thousand, rests on a number of assumptions we should name honestly.
The first assumption is that AI models become more efficient. So far that has held true: foundation models get faster, cheaper and better with every generation. But there is no law of nature that says this pace continues. If the efficiency gains flatten, the cost model becomes fragile.
We run our models on our own GPU infrastructure, through a Dutch sister company with dedicated Kubernetes and GPU capacity. That means we have no variable token costs, and that customer data and findings stay within our own server cluster. For many organisations in our target group, that last point may well be the most important argument. But owning GPUs shifts the problem, it does not eliminate it. The question then is not “what does this test cost?”, but “how many tests can we run at the same time?”. The limit is capacity, not budget per test.
The second assumption is that human involvement can stay minimal. If the L3 validation, the human review by a senior pentester, takes up a significant share of the turnaround time, then a ninety percent cost reduction cannot be reached. The maths is simple: if validation costs thirty percent of the time of a conventional test, then your maximum saving is seventy percent, not ninety. The CCV mark makes that calculation sharper still: fifty percent human time means a ceiling of fifty percent saving, not ninety.
There are already players in this field
Guardian360 is not the first to try to automate penetration testing. Horizon3.ai offers autonomous pentests for internal networks and Active Directory environments with NodeZero. Pentera positions itself as automated security validation and reports more than a hundred million dollars in annual recurring revenue. Hadrian focuses on the external attack surface and launched an agentic AI pentest product in March 2026.
The question is fair: why build something yourself if established platforms already exist? Our answer is threefold. First, these platforms are closed source. Our agent and methodology will be published as open source after the project, so the community can test, improve and adapt them. Second, they are predominantly American, while European regulation, in particular NIS2, DORA and sector-specific standards such as NEN 7510 and the BIO framework, requires its own scenario libraries. Third, they are expensive: the usual licensing models are designed for enterprises, not for small and medium-sized businesses.
What we are learning
We are about two months in. That is too short for conclusions, but long enough to share the first insights.
The first insight is that the agent is strong where humans are weak: in persistence. A human pentester gets less sharp after hours. The agent does not. It works through every aspect of an environment systematically, including the boring parts. That yields findings a human would have missed, not because the human cannot find them, but simply because a human is not a machine.
The second insight is that the gap between “thinking up a scenario” and “running that scenario in a real environment” is larger than expected. An attack scenario that is logical on paper gets stuck in practice on session management, MFA flows, rate limiting or application-specific authentication patterns. Every environment has peculiarities the agent does not know until it meets them. That means the scenario libraries must contain not only attack patterns, but also the defensive patterns the agent must recognise and either work around or report.
The third insight is that per-sector scenario libraries are not a luxury but a necessity. A pentest for a healthcare provider that has to meet NEN 7510 asks for different scenarios than a test for a financial institution under DORA. Generic scenarios produce generic findings, and generic findings are not usable for a CISO.
The fourth insight, and perhaps the most confronting, is that the reaction of experienced pentesters to AI-generated findings is a measure of value we cannot automate. The agent’s reconnaissance and first triage are usable; there it saves the human hours. But the degree to which the human reviewer has to correct, reclassify or supplement findings afterwards ultimately determines whether the model is profitable. We now measure that feedback loop systematically, and the outcomes are still too early to publish.
A broader European story
As noted, Guardian360 is not the only organisation working on this. The CYSSDE programme selected thirteen projects in Open Call 3 from nine European countries, together planning more than 534 security assessments. From OT security assessments in Poland and Germany, through threat-led pentesting in Belgium, to sovereign AI pentesting in Italy.
That is no coincidence. The EU is deliberately investing in European cybersecurity capacity, through the European Cybersecurity Competence Centre and programmes like CYSSDE. The thinking behind it is that if we want European organisations to be resilient, we should not depend on non-European tools and suppliers. Our open-source commitment fits that same line: it is a contribution to the European cybersecurity commons, not just a product.
The honest answer
Can penetration testing be ninety percent cheaper? The honest answer is: we do not know yet. We believe AI can take over a large part of the repetitive work and that costs can fall substantially as a result. But “substantially” is not the same as “ninety percent”, and the pitfalls we are running into, from business logic the agent does not understand to quality standards like the CCV mark that enforce human involvement, are cause for modesty.
What we do know is that the current situation is untenable. Organisations that fall under NIS2 and cannot afford a pentest are at risk. Organisations that have a cheap automated scan run and think they got a penetration test run a different risk: false security. A pentest that gives organisations a false sense of security is perhaps even worse than no pentest.
The solution is not in AI alone, and not in people alone. It is in a combination that is honest about what each part can and cannot do. The agent scans, the human judges. The agent finds patterns, the human sees chains. The agent scales, the human guards the boundary.
We publish our results, our agent and our methodology. Not because we have the answer, but because we believe the answer comes sooner if more people are looking.
Sources
- CCV Keurmerk Pentesten, Certification Scheme version 2.0 (January 2024), https://hetccv.nl/keurmerken/cybersecurity/pentest/
- CYSSDE Open Call 3, selected projects, https://cyssde.eu/cyssde-activities/open-call-3/
- Guardian360 AI Pentesting, beta programme, https://guardian360.net/platform/ai-pentesting/
- NIS2 directive (EU 2022/2555), Article 21: cybersecurity risk-management measures
- Security Magazine (2026), “Would you trust an AI pentester to work solo?”, https://www.securitymagazine.com/blogs/14-security-blog/post/102154
- NHI Management Group (2026), “AI pentesting replaces repetition, not expert judgment”, https://nhimg.org/articles/ai-pentesting-replaces-repetition-not-expert-judgment/
- The Next Web (2026), “The 10 best autonomous pentesting tools, ranked by proof of exploit”, https://thenextweb.com/news/best-autonomous-pentesting-tools-proof-exploit-2026