Longtime web security researcher James Kettle presented findings at Black Hat in Las Vegas on Wednesday that put agentic AI's offensive security abilities in perspective. His conclusion: AI alone is minimally capable of devising new attack paths, but with human help it becomes a powerful research partner.

Kettle set out to test whether agentic AI could develop novel, abstract hacking methods from concept to practical attack. Early on, he found the models tried to pass off existing research as original by surfacing obscure findings that were hard to verify. To regain control, he scoped experiments to his own web security specialty, where he could identify false results. He also trained models on his own research methodology to see how far they could extrapolate.

Over months of experiments starting in September 2025, using the latest Anthropic and OpenAI models, Kettle saw the systems produce findings far faster than he could. He reported notable discoveries roughly every two days, and in a few months they found more proven examples of certain vulnerabilities than he expected to find in years. The workload became overwhelming and pushed him to automate more analysis.

The most significant result came from human/AI collaboration: a new vulnerability area Kettle calls Shared-Parser Confusion. It stems from web servers using shared code to process both untrusted requests and trusted responses. The AI generated the hypothesis from real, proven findings, and Kettle evaluated and confirmed it. He says he would not have found it alone, but the AI could not prove it alone either. Kettle stressed the finding shows how AI can contribute to both defensive and offensive security work, while the limits are still rarely discussed.

Source: Wired