Agentic AI has made it faster and simpler to find software flaws or turn them into weapons. At the Black Hat security conference in Las Vegas on Wednesday, James Kettle presented research showing that while AI cannot yet invent new hacking paths entirely on its own, it becomes an extremely powerful partner when guided by humans.
Shared-Parser Confusion
Kettle, a veteran web security researcher, identified a new vulnerability he calls Shared-Parser Confusion. This issue arises because web servers often use shared code to handle both incoming requests and outgoing responses. Requests are inherently untrusted, but responses are usually treated as safe. Kettle told WIRED that this gap creates a massive attack surface.
“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 said.
The research began in September 2025. Kettle tested Anthropic and OpenAI models to see if they could conduct theoretical security work. He found that AI systems often tried to present findings on extremely obscure topics as original work, making them hard to verify. To avoid this, Kettle restricted the AI to his specific area of expertise. He knew the material and could not be tricked. By training models on his own research methods, he could test how far the systems could extrapolate.
Kettle stated he wanted to push AI to the absolute limit to see where it fails and where human input is required. He noted that few people discuss these limits because there are no incentives to admit where a system breaks down.
A productive feedback loop
As the experiments progressed and more powerful models were released, the systems generated findings at a rate far exceeding Kettle’s own output. This created a productive research feedback loop.
Kettle described the process as notable, with new findings appearing every two days without him needing to log in. The volume of leads caused anxiety and a fear of missing out, forcing him to automate more analysis.
The AI found more examples of known vulnerabilities in a few months than Kettle likely could have in years. It also succeeded in identifying a novel class of bugs, though the specific Shared-Parser Confusion finding was extremely rare and not exploitable in the single target available.
Kettle emphasised that the discovery was significant because it showed how AI and humans can work together for both defensive and offensive hacking.
“It wasn’t able to prove this itself, but it analyzed some real, proven findings and came up with the hypothesis, and I evaluated it and confirmed it,” Kettle said. “That’s probably going to be the discovery that has the biggest long-term impact. It couldn’t do that on its own, but I would never have found that on my own for sure.”
Even with the single line of documentation available, Kettle admitted he would not have seen the flaw alone. Together, they managed to find it.
