Jason Haddix: Stop fearing AI pentesting Jason Haddix, former CISO of Ubisoft and author of The Bug Hunter's Methodology, predicts that 90% of penetration tests will be performed by AI in the near future, arguing that defenders must 'fight AI with AI' to keep pace with AI-driven attacks. According to Aikido's State of AI in Pentesting 2026 report, 79% of 400 CISOs and engineering leaders worry about vulnerabilities slipping between scheduled tests, and 48% say findings are outdated by arrival. Haddix says AI will eliminate 'scanners in disguise' and checkbox pentests, but stresses that effective AI pentesting requires baking in human methodology, as raw AI models 'suck' at finding bugs without it. This post is based on Mackenzie's conversation with Jason Haddix on The Secure Disclosure podcast . Listen to the full episode or watch below. Jason Haddix has topped the Bugcrowd leaderboards, run security as CISO of Ubisoft, and wrote The Bug Hunter's Methodology https://github.com/jhaddix/tbhm , the playbook a lot of working pentesters learned from, so his read on the direction of the field is worth listening to. He believes that “90% of pen tests will be done by AI” in the near future. As AI is lowering the barrier of entry to cyber attacks and drastically increasing the volume and speed of malicious operations , the only way defenders can keep up is to put AI to work on their own side /blog/hacker-superpower-ai . As Haddix says, “We're going to have to fight AI with AI.” Naturally, this scares human pentesters are their jobs evaporating? Haddix doesn't think AI pentesting is an extinction event for humans, though it is something that will inevitably impact their existing workflow. AI solves pentesting's coverage problem Manual pentesting is not only cost-prohibitive, its typical twice-a-year cadence is not adequate for how fast teams are deploying AI-generated code. In Aikido's State of AI in Pentesting 2026 /state-of-ai-pentesting report, 79% of the 400 CISOs and engineering leaders surveyed said they worry about vulnerabilities slipping in between scheduled tests, and 48% said findings are already outdated by the time they arrive. {{cta}} "The biggest problem we've always had is scale," Haddix says. If you run security for a large organization, the math meant you were never going to manually pentest everything your team shipped. Covering all of it with manual testing "would have been a billion dollars in pentesting." So you tested biannually and hoped for the best. "We have prayed for scale before," Haddix says, "and now they're just railing against it." His read on the people leading that resistance is that they "want to stand back and sling mud when they should be jumping in and helping people build." What actually dies "Scanners in disguise," is what Haddix calls shops that run open-source scanners and repackage the results. These work differently from real pentests, and return much worse results, which is why they will be the first to go. The other casualty is the checkbox pentest a customer buys just to satisfy an auditor. Performing these tests has always been the first job a tester takes to break in, and the people in it rarely want to stay. The people doing it would rather be learning real exploits and building their own methodology. AI clears out the drudge rung. Haddix is honest about the catch, though. If AI takes the foot-in-the-door work, where does the next generation of testers get their start? "I don't have a great answer for that," he admits. "We're still thinking about it." AI pentesting only works with human methodology baked in Point Claude Code at a target and tell it to find bugs, and "it's going to suck. Even the best model, max thinking, doesn't matter," says Haddix. "The power comes when you bake in your methodology. Take Haddix's recon process, which has 18 steps for mapping the sites a big organization actually owns. Two or three of them used to need a human's eyes and judgment. A company like Nintendo has registered hundreds of thousands of domains, most of them parked, and deciding which ones host a real target was hard to automate. Now an agent reads the page and makes that call itself. Across the industry, "the smarter ones are hiring pentesters who have unique methodologies and codifying it, putting it in prompting," Haddix says, and "the smartest ones are training synthetic data sets" to keep that reasoning accurate. And once it's baked in, AI does the creative, chained work that people assume only humans can. In one case from Aikido's own benchmark /reports/autonomous-vs-manual-pentesting-benchmark , agents testing a document-signing app worked through its logic and proved that e-signatures could be forged, all in roughly nine hours. The senior human team spent about two weeks on the same app, returned one XSS and one SSRF, with seven of their nine findings being routine hardening checks, and never caught the forgery. Giving a pentester access to an application's source code results in a deeper test. Most human-run pentests run grey or black box, limiting the depth of a test. Across more than 1,000 AI pentests, Aikido found https://www.aikido.dev/blog/ai-pentesting-buyers-guide that whitebox testing uncovered 7x more vulnerabilities while requiring fewer attempts than greybox testing alone. Humans generally don’t do whitebox testing, because checking every path in the code isn’t feasible. An AI pentest with whitebox testing is able to coordinate a fleet of agents to check every line of the code, and connect it back to the live environment. While you can bake methodology in, humans miss things too. In Aikido's State of AI in Pentesting 2026 /state-of-ai-pentesting report, half of the surveyed 400 CISOs and engineering leaders say logic flaws, broken access controls, and multi-step bugs get missed always or often in manual testing, and that jumped to 92% for teams shipping daily or faster. This shows that while the human methodology is well grounded, it's by no means perfect, and AI helps cover what a manual pass would miss. Aikido AI Pentesting is built on expertise Aikido Security / built AI Pentesting /attack/aipentest with the expertise that Haddix describes, and independent testing backs it up. Doyensec benchmarked Aikido /reports/aikido-vs-xbow on two open-source apps drawn at random from 442 candidates, with a senior researcher validating every finding by hand. Aikido surfaced 49 verified vulnerabilities, 9 of them high or critical, at a 4% false-positive rate, and was running self-serve on both apps in under 20 minutes. {{pentest}}