AI continues to play a role in security—and for those on the frontlines, the potential of automation is tempered by the need for human oversight.
That’s most starkly seen in research from pen testing company Cobalt published on November 17. The Autonomous Pentesting Pulse Report found that, while 98% of surveyed pen testers said they were using AI “at least a little,” only 2% reported being comfortable with allowing AI to autonomously “chain a known exploit into a multi-step attack without supervision.” (The report collected data from September 2026 in partnership with Omdia.)
Options, open. Cobalt CTO Gunter Ollmann told IT Brew that’s likely because of the capability expansion offered by AI, which allows users to shift their attention from menial tasks to more creative solutions to problems.
“If I can use AI and AI tools to make my job easier so I can focus on more interesting things and more interesting challenges—then [it’s] no surprise that every one of these pen testers and would-be hackers are using these technologies,” Ollmann said.
He cautioned that the technology should be considered untrustworthy because of its need for guardrails. That explains the 2% who believe it can act alone, who Ollmann described as “really fully trusting, and believe that it can work through that process; [they] probably aren’t the best pen testers.”
Using AI. Pen testers have a complicated, difficult job. They are often the first line of defense in finding vulnerabilities in systems, and work outside normal security parameters. AI has been used for years to assist in pen testing; where the technology can provide use, Ollmann said, is in catching the little things, which require large investments in time and capacity.
Catching the little things is key, and requires a large investment in time and capacity.
“Leveraging and using AI can help you perform many of these activities, but the real security threats tend to be in those narrow scope gaps,” Ollmann said. “From my side, where we’re building and training AI systems, the data that’s necessary to train these super efficient fully autonomous pen testing tools is still very large.”