Will AI replace penetration testers?

Time studies are rare in this field, but the shape is not controversial among practitioners: a large share of engagement hours went to enumeration, running known techniques against a surface, and writing up findings. A smaller share went to the work people describe when asked what the job is, which is figuring out what matters and why.

That distribution is why automation reads as threatening. If most of the hours automate, it looks like most of the job does. But hours and value were never the same distribution, and the hours that automate are disproportionately the ones that were being rationed rather than chosen.

The clearer way to put it is that automation removes the constraint that forced testers to sample. What was scarce was attention, and attention is the input that just got cheaper.

When findings are cheap to generate, the scarce thing becomes the assurance that a finding is real. That is a different skill from producing it, and it is harder to fake.

OWASP names misinformation as LLM09:2025, which in this context describes a specific and common artefact: a finding that is fluent, plausibly evidenced, and wrong. Distinguishing those from real results requires someone who can reproduce the path by hand and recognise when the evidence is an artefact of the testing setup rather than the estate.

This is also where accountability sits. A report is read by people who will spend money because of it, and somewhere in the chain a person has to be answerable for the claim that the path exists. No amount of automation moves that, which is why the signature at the bottom of a report is not ceremonial.

Every finding is reproduced and verified by a Cryvanta operator before it reaches you. No unranked scanner output, no false-positive triage left on your desk.

This is the part of the question worth taking seriously, and it gets the least attention because it is inconvenient for everyone.

Practitioners did not learn judgement in a course. They learned it by doing thousands of hours of enumeration and technique-running, developing intuition about what a surface looks like when something is wrong. Those hours are exactly what automates first, which removes the ladder while leaving the roof.

There is no established answer. Deliberate practice on automated output is one route: reviewing machine findings and being graded on which ones were wrong builds the same discrimination faster, if someone senior invests the time. What does not work is assuming judgement will appear in people who have only ever supervised, because supervision without ground truth is how confident wrong answers get approved.

For a buyer this is not an abstract concern. The quality of the verification you are paying for in five years depends on whether the providers you use today are training anyone.

The amount of testing you need did not fall. If anything the opposite, since coverage that was previously unaffordable is now in reach and the estate has not become simpler.

What changes is what a good deliverable looks like. Volume of findings was always a weak quality signal and is now close to meaningless, because generating findings is the cheap part. The signals worth reading are whether findings reproduce, whether the impact judgement shows knowledge of your business, and whether the scope decision is written down with its reasoning.

Under NIS2 that alignment is convenient. Article 21 asks for policies to assess whether measures are effective, and an assessment made of reproducible cases is easier to stand behind in front of a supervisory authority than a long list nobody has confirmed.