AI Uncovers New Web Vulnerability Class at Black Hat
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- James Kettle presented findings at the Black Hat security conference in Las Vegas on Wednesday, concluding that agentic AI is 'minimally capable but extremely limited' in autonomously devising new attack paths.
- Kettle discovered an entirely new vulnerability class dubbed 'Shared-Parser Confusion,' stemming from an AI-generated hypothesis that web servers use shared code to process both untrusted requests and trusted responses—a 'major attack surface.'
- The experiments began in September 2025 using Anthropic's and OpenAI's latest models at the time, after Kettle narrowed tests to his own web security expertise because the systems initially tried to disguise existing research as original findings.
- As Kettle fed the models more methodological data and refined parameters, the systems produced notable findings every two days without his input—creating what he called a 'productive research feedback loop' that outpaced his manual pace.
- Kettle confirmed the Shared-Parser Confusion finding required collaboration: the AI generated the hypothesis from analyzing proven findings, but Kettle evaluated and verified it himself, noting 'I would never have found that on my own for sure.'
Why it matters: This is rare empirical evidence that AI's strongest current cybersecurity contribution is augmenting expert humans, not replacing them—a distinction that matters for both defensive teams scaling vulnerability discovery and for offensive risk modeling where novel exploit classes still demand expert verification before weaponization.
Ask SkimNews
