Can AI do novel security research - Meet the HTTP Terminator

The talk asks whether an autonomous system can invent new HTTP desync attack techniques and then use them against live websites. Desync research was chosen because each candidate can be proven on a real target instead of remaining a theory. The pipeline has four stages: generating hypotheses from fragments of RFC text, testing candidate triggers across roughly 30,000 bug bounty sites using plain code rather than AI, weaponizing confirmed hits, and feeding each discovery back in to seed further discoveries.

The generator produced about 30,000 unique vectors from 138 RFCs, and several of them caused desync on production systems, including banks and an airport. Early attempts showed that coding agents struggled to build working exploits, so success checks were moved into deterministic code and each stage received only evidence from the previous one. A technique that omits one byte from a second request removed the race condition behind stacked responses. Later cascades produced a zero-day in Apache Traffic Server and a proposed class called shared parser confusion, where servers reuse response-processing code for requests.

The main takeaway is that automated research can surface new issues, but the largest gains came when a human guided the cascade stages. Evaluation design decides whether results are signal or noise. The speaker suggests choosing topics where automated checking is hard if you want to avoid collisions with AI-driven researchers. The full source code of the system has been published.