Longtime web security researcher James Kettle presented findings at the Black Hat security conference in Las Vegas on Wednesday that address a pressing question: can agentic AI develop novel, abstract hacking methods from concept through to practical attacks? According to WIRED reporter Lily Hay Newman, Kettle concluded that AI is "perhaps minimally capable but extremely limited" when it must devise new attack paths fully autonomously, but when paired with human guidance, it becomes an "extremely powerful partner" in conceptualizing and uncovering new hacking strategies.
Agentic AI's Rapid Rise in Cybersecurity
Agentic AI — systems that can take goal-directed actions — has permanently changed cybersecurity by making it quicker and easier to discover vulnerabilities in software and fix them, or develop so-called exploits to weaponize them, WIRED reported. Kettle's experiments, which began in September 2025, used Anthropic and OpenAI's latest models available at the time. He deliberately scoped his tests to his own web security expertise and synthesized his research methodology to train the models, guarding against AI systems passing off existing research as original by returning findings about esoteric topics that were difficult to vet.
Speed of Discovery: Notable Findings Every Two Days
Kettle told WIRED that the systems produced notable findings roughly every two days "without me even logging into the system," generating so many research leads that he felt FOMO about not exploring them all and was forced to automate more analysis. Over a few months, the AI found more proven examples of certain vulnerabilities than Kettle could likely find in a few years, he said. As he provided more methodological data and refined parameters, the systems' outputs created what he described as a productive research feedback loop.
Shared-Parser Confusion: A New Vulnerability Class
The most significant output was an entirely new area of potential vulnerability dubbed Shared-Parser Confusion, stemming from an AI revelation about web servers using shared code to process both requests and responses. "This is an absolutely massive deal, because if you think about it, requests to a website are completely untrusted, they could be anything, but responses are trusted," Kettle told WIRED ahead of his talk. "So this is a major attack surface and potentially spills into a lot of different attack types."
Where AI Still Falls Short
Despite that conceptual breakthrough, Kettle emphasized that the finding related to an extremely rare type of bug and was not actually exploitable in the one vulnerable target available. He argued the limits of AI deserve more attention: "There are still very few people talking about where the limits are, especially in the security space, because there aren't incentives to talk about that angle. Everyone wants to be seen as AI native, not talk about where their system falls apart completely."
| Aspect | Outcome | Attribution |
|---|---|---|
| Fully autonomous novel attack design | Minimally capable, extremely limited | Kettle's presentation at Black Hat |
| AI paired with human guidance | Extremely powerful partner in new strategies | Kettle's experiments, September 2025 onward |
| Pace of findings | Notable findings every ~2 days; more vulnerabilities in months than years | Kettle told WIRED |
| Shared-Parser Confusion | New vulnerability class discovered, but not exploitable in one target | WIRED coverage |
Implications for Enterprise Security Leaders
For CTOs and security teams, the research illustrates that agentic AI can accelerate vulnerability discovery and remediation, but outputs still require human verification and contextual insight. Major AI organizations have disclosed real-world examples of rogue AI hacking, WIRED reported, making the limits of autonomous AI a business-critical concern. Kettle's methodology — pairing his own research methodology with model training — mirrors a practical workflow for enterprise teams: let AI generate leads at scale, then put experienced analysts in the loop to validate, refine, and decide what to act on. The Black Hat findings show the speed gains are real, but so is the persistent need for human judgment in security research.
Sources:
