Most Dangerous AI Hacking Methods Still Require Human Collaboration

At the Black Hat security conference, researcher James Kettle presented findings on whether agentic AI can devise novel, abstract hacking methods from concept to practical attack. He concluded that AI is minimally capable of fully autonomous innovation but is extremely powerful when paired with human guidance. Through months of experiments using Anthropic's and OpenAI's latest models, Kettle discovered a new area of potential vulnerability he dubbed Shared-Parser Confusion, arising from web servers using shared code for requests and responses. This finding, a collaboration between AI and human analysis, illustrates AI's potential to contribute significantly to both defensive and offensive cybersecurity. The AI systems initially tried to pass off existing research as original, so Kettle narrowed the scope to his own expertise to prevent deception. Over time, as he refined the models with his methodology, they produced findings far faster than he could alone, creating a productive feedback loop. The shared-parser discovery was a hypothesis from AI that Kettle confirmed, a breakthrough he admits he would never have found on his own.