Bugcrowd Braden Russell TechSpective Podcast

The Hidden Cost of AI Pen Testing

Years ago, my mother-in-law had a computer that seemed to catch every virus making the rounds. She had two or three different anti-malware products running at the same time, and she still got infected on a regular basis. I wasn’t running anything, and I wasn’t getting infected. Stacking more security tools on that machine didn’t make her any safer, but it did give her more pop-ups.

AI-powered penetration testing has a version of that same trap built into it. Vendors keep rolling out agentic pen testing tools, and they’re getting very good at finding things. That’s where the cost nobody puts on the price sheet starts to pile up.

Why Security Teams Want AI Pen Testing

Companies used to be able to make a rough guess at how fast the attack surface would change, because a certain number of developers could only write so much code. AI coding tools have thrown that math out the window. Applications now change faster than security teams can keep up, and the traditional pen test hasn’t kept pace.

Most organizations test their crown jewel applications once or twice a year. That leaves all the releases in between, plus the other 400 or 500 applications and assets that never got crown jewel treatment. Historically, that long tail got a scanner and a crossed finger or two.

Braden Russell, CTO at Bugcrowd, told me that AI agents have raised the floor on what automation can detect by orders of magnitude. Agents can run broad testing across that whole long tail and return deeper, higher-quality results than legacy scanners ever did. They also don’t sleep, and they can keep threat models current in a CI/CD world where a new build ships every day.

Somebody Has to Validate the Findings

All that breadth comes with a lot of noise, though. Most security pros have chased a suspicious file down a rabbit hole, pushing it to VirusTotal and digging through it, only to find out it was harmless the whole time. Multiply that by thousands of AI-generated findings, and you get a sense of the problem.

Russell, who has also written about the hidden cost of agentic pentesting, pointed out something the market doesn’t talk about much. Many of these platforms proved themselves by climbing leaderboards on bug bounty and vulnerability disclosure programs, and those programs already had a triage layer between the agent and the customer. When a company just licenses the AI tool on its own, that layer goes away, and the job lands on the security team. Instead of a force multiplier, Russell said, “it’s just giving me a whole lot more work.”

His suggestion is to ask a vendor what happens between an agent finding something and your team seeing it. If the answer amounts to dumping raw findings in your lap, then by his description, “you’re paying to be the validation layer for that platform.” Bugcrowd’s approach with Savant Pathseeker, its agentic pentesting product, is to run findings through the same human and automated triage process it uses for its bug bounty programs.

The human-in-the-loop argument has a flip side, too. If you could validate every AI result by hand, you wouldn’t need the AI in the first place. The USDA doesn’t inspect every pound of ground beef. It spot-checks samples and makes a call on the batch. Validating AI findings is going to look a lot more like that, so somebody has to decide what gets checked and how.

Building It Yourself Costs More Than the Model

Some security teams figure they have access to the same frontier models as everyone else, so why not wire up a harness and build their own? Russell says teams learn a lot by doing it. The model is the easy part, though.

Somebody has to manage credentials at multiple permission levels, because the unauthenticated attack surface is a small slice of the whole thing. Logging in through an SSO or identity provider can send the agent out to someone else’s infrastructure, and suddenly the agent may be testing systems you don’t own. Staging and production are rarely as identical as people like to think, and plenty of breach-causing misconfigurations only exist in production.

Maintenance adds up, too. Every new model release means checking whether the results are still as good as they were last time. Running a frontier model for recon and report writing also burns money on routine work that a cheaper model could handle. “There’s a whole lot more to the car around the engine,” Russell said.

Where Humans Still Fit

Agents still struggle with intent. An agent can tell you an endpoint accepts negative numbers, and it can tell you the app issues refunds, but it may not connect those two facts into a refund for something nobody bought. Regulated environments and audits also still want a person’s name on the report.

I expect that balance to shift over time. The first time I used adaptive cruise control, I kept my hands right next to the wheel, and now I’m perfectly happy riding in a Waymo. AI pen testing will earn more trust as it proves itself, but humans will keep finding novel attacks first and teaching them to the AI later.

Russell considers AI the best tool available right now for offensive testing, and testers who ignore it are putting themselves at a disadvantage. The license fee just isn’t the whole bill, though, and the part missing from the invoice usually lands on the security team’s plate.

Check out the full TechSpective Podcast episode here:

Scroll to Top