Sponsored by PlexTrac.

I'd bet most pentest teams have had some version of this meeting in 2026. One of the testers spent a weekend with Claude Code, and on Monday there's a working demo of an internal reporting tool. It pulls in Burp output, drafts the findings, and spits out a clean PDF. Everyone gets a little excited, because nobody got into offensive security to fight with Word templates.

I'm genuinely sympathetic. I build a ridiculous amount of my own tooling with AI now. Standing up a prototype has never been cheaper, and it's fun.

So I wanted to actually run the numbers on what happens after the demo. PlexTrac's founder published his version of this math and I'll get to it, but I also wanted data from outside the vendor's own walls: breach data, remediation data, labor data, and what the research says about the code AI is writing for all of us right now.

Short version: the build got cheap, and the cost of owning it stayed right where it was.

The report is where remediation starts

Quick reminder of why any of this matters. Verizon's 2026 DBIR found vulnerability exploitation is now the top initial access vector, behind 31% of breaches, up from 20% the year before. Credential abuse, the long-running champ, fell to 13%.

That puts the spotlight on how fast vulnerabilities actually get fixed, and the same report says we're getting slower. Organizations fully remediated only 26% of the vulnerabilities on CISA's Known Exploited list in 2025, down from 38% the year before. The median time to fix went from 32 days to 43. Roughly 60 to 70% of KEV bugs were still open a week after discovery, while the KEV list itself grew 50%.

So the bugs attackers are actively using are piling up faster, and we're closing a smaller share of them, more slowly.

I've spent a lot of my career on the generating and receiving end of pentest reports. When you run vuln management across big codebases, the report is where your work starts. Somebody has to turn every finding into a ticket, route it to the right team, argue about severity, chase the fix, and get it retested. If the report shows up as a static PDF, a human spends a day copying findings into Jira before any of that begins.

❝

Every day a finding sits in a PDF waiting for someone to retype it is a day added to that half-life.

So whatever reporting tool you run sits right on the hinge between finding a bug and fixing it. It also holds the most sensitive data your team touches. That's a lot of responsibility for a weekend project.

Reporting still eats the engagement

Now the testers' side. Ask anyone who has written a pentest report where the engagement time goes and you'll get the same answer. Reporting is the biggest chunk of non-testing time, and a shocking amount of it is pasting the same writeup you used last month, reformatting scanner output, and fighting a template. PlexTrac reports customers cut reporting time by up to 75%. Apply the usual salt to any vendor's number, including the sponsor's, but the direction is right.

LLMs are also very good at drafting prose from notes, which is the other half of the problem.

That's exactly why the build pitch is so tempting. The pain is real and the tool to fix it is sitting in every tester's terminal.

Weekend one is the cheap part

PlexTrac founder and CTO Dan DeCloss wrote up a build vs. buy breakdown that's worth reading. His argument, which I agree with, is that AI speeds up the build and leaves the upkeep right where it was. Long term ownership, in his words, is "where the real cost lives."

His model adds up the boring stuff: hosting and patching, parser upkeep across your scanners, supporting your own testers, template changes for clients, reviewing AI output, and securing the tool itself. It lands around 900 hours a year, roughly half an engineer, just to keep the tool where it already is.

Just like open source, it’s free like a puppy, not free like beer.

I wanted a dollar figure on that, so here's my napkin math using public labor data:

  • BLS puts the median software developer at $135,980 (May 2025).

  • For full-time private industry workers, benefits are 31.5% of total compensation (June 2026).

  • That's a loaded cost of roughly $198,500 a year, or about $110 an hour.

  • Half an FTE comes out to about $99,000 a year, before hosting, LLM API spend, or any new features.

Call it $300K over three years for a tool that, by design, never gets better than the day it shipped. And that assumes a median developer maintains it. The person who understands pentest workflow well enough to own this thing is usually a senior tester whose time is worth a lot more on an engagement.

❝

Half an engineer, every year, to keep the tool exactly where it is.

The broader industry data says that estimate is probably generous. Stripe's Developer Coefficient research found developers spend about 42% of their time on technical debt and maintenance. Chainguard's 2026 Engineering Reality Report puts it far higher, with engineers saying only 16% of their week goes to building new things. Your homegrown reporting tool joins that queue the moment it ships.

Here's what that queue looks like in real life:

  • Tenable or Burp tweaks an export and your importer starts quietly dropping fields.

  • Your findings library drifts out of date because nobody actually owns it.

  • A big client wants their own report layout. Then another one does.

  • The tool falls over mid engagement, and the only person who understands it is on a plane to a client site.

  • Somebody has to secure a system holding every client's unpatched vulns. Ideally someone other than the tester who wrote it on a Saturday.

The detail from DeCloss's post that stuck with me: a pentest consultancy that was a PlexTrac customer left for a few years to build their own platform, then came back this year. That's a real team that made the build call in good faith and decided it wasn't worth carrying.

AI-written code makes year two heavier

Here's the part I think gets skipped in these conversations. The weekend prototype is mostly AI-written, and the research on AI-written code says the maintenance load is growing.

GitClear's June 2026 study looked at 623 million code changes from 2023 to 2026. Code duplication is up 81%. Refactoring is down 70%. Updates to code older than a year fell 74%. And the one I'd circle in red for a reporting tool: code that catches errors without dealing with their cause is up 47%.

Think about what error masking means in this specific tool. A parser hits a field it doesn't expect, swallows the exception, and moves on. The run finishes clean, and a finding just never makes it into the client deliverable, and nobody notices until the client's own scanner does.

GitClear 2026 maintainability metrics

Then there's security. Veracode's 2026 GenAI Code Security Report tested more than 100 models and found an average security pass rate of 56%, basically flat from last year, while syntax pass rates sit near 100%. Coding-specific models scored no better than general-purpose ones. The best model in the test still failed about a third of the security tasks.

So the tool that holds every client's unpatched vulnerabilities was largely written by a model that, on Veracode's benchmark, writes vulnerable code a little under half the time when nobody asks it to be careful. Somebody has to pentest the pentest platform. Ideally before a client's security team asks for the results.

The agent adds three line items

I'll give the agent crowd their due. Having a model draft a finding from your notes and evidence is useful, and it's going to keep getting better. It also adds three things to your maintenance list that didn't exist on anyone's cost model two years ago.

Where the data goes. Harmonic Security analyzed tens of thousands of enterprise prompts and found 8.5% contained sensitive data. Security data made up about 7% of that, with pentest results and network configs called out as examples, and a big share of usage was on free tiers. If your homegrown agent ships client findings to a model provider, you now own the answers on which provider, what gets retained, and who can see the prompts. Your clients' security questionnaires are going to ask.

Who signs the report. Models can be confidently wrong, and your tester's name is still on every finding. Someone reviews every AI draft before it reaches a client, which is peer review with extra steps.

The model moves underneath you. Anthropic gives at least 60 days' notice before retiring a public model. On September 30 it told developers that Claude Sonnet 4.5 retires November 30, 2026. If your agent's prompts were tuned against that model this spring, someone on your team is re-validating finding output this month. Every provider has some version of this cycle, and the cadence keeps getting faster.

What teams are actually scared of losing

When teams argue for building, the worry usually comes down to three things. They have a house report style their clients recognize. They have years of findings writeups that represent the team's real expertise. And they have a toolchain of scanners and ticketing they don't want to rip out.

Those are all legit, and they're the same three things a decent platform should hand back to you. On PlexTrac specifically:

  • Your format. Templates are customizable, and deliverables go out branded as yours, either as a document or through a white-labeled client portal where clients watch findings update live.

  • Your expertise. You load your own writeups into their content library and reuse them across engagements, alongside a stock library of 25,000+ writeups. Early Warning Services' offensive security lead said loading 170-plus of their own writeups got them to about 90% of report content pre-built.

  • Your tools. Imports from the usual suspects (Burp, Nessus and Tenable, Nuclei, Qualys, NodeZero, HackerOne, Bugcrowd, plus CSV), two-way Jira sync, ServiceNow, Slack and Teams, and an open API if you need a custom import.

  • The after part. Remediation tracking and retesting live next to the report. Given those DBIR remediation numbers.

Their pricing also doesn't scale with your engagement count, which matters a lot if you're a consultancy whose whole business plan is "more engagements."

When building is the right call

To DeCloss's credit, his post makes the case for building too, and I agree with it. If you work in an air-gapped environment with classification and compliance requirements nobody else shares, you might have to build. If owning the roadmap matters more to you than the hours, build. And if you're a small shop doing a handful of reports a year, you don't need a platform yet. Word and Excel are fine.

If your team loves building internal tools and you have the headcount to staff it like a product, go for it. Just actually staff it like a product, with an owner, a backlog, and someone who pentests it. The failure mode is a tool that becomes critical while still being owned by whoever had a free weekend.

Where this leaves us

AI answered the "can we build it" question for almost everyone. What's left is whether you want to own a piece of production software that holds your most sensitive client data, sits between every finding and its fix, and was mostly written by a model that fails security tests close to half the time.

Some teams should say yes to that, and they know who they are. For everyone else, the math is pretty simple.

Build the weekend prototype. You'll learn a lot about your own workflow, and that's useful whatever you decide. Then, before a client ever logs into it, put a name next to it. Figure out how many hours of testing that person gives up every year to keep it alive, and price that against PlexTrac.