James Kettle Finds Human Guidance Still Powers AI Hacking Research
PortSwigger research director James Kettle presented an August 5 Black Hat study of whether AI agents can devise novel security research. WIRED reports that fully autonomous systems remained limited, but targeted human guidance helped produce a new web-security hypothesis called Shared-Parser Confusion. The result points to a powerful human-AI research loop, not an autonomous replacement for expert validation.
James Kettle, director of research at PortSwigger, presented an August 5 Black Hat briefing on whether AI systems can produce genuinely novel security research. The official conference program lists the session as “Can AI Do Novel Security Research? Meet the HTTP Terminator.”
According to WIRED's reporting from the conference, Kettle spent months testing models from Anthropic and OpenAI against web-security research tasks. Fully autonomous work remained constrained: systems sometimes resurfaced obscure existing research as if it were new, and they struggled to turn rare findings into practical exploits without expert judgment.
A productive loop, not an autonomous researcher
Kettle narrowed the experiments to an area where he could independently check originality and correctness. He supplied more of his research method, refined the tasks and evaluated promising outputs. WIRED reports that this human-guided process generated useful leads faster than Kettle expected.
The most consequential result was a hypothesis Kettle calls Shared-Parser Confusion. It focuses on web servers that reuse parsing code for both untrusted requests and trusted responses, creating a potential attack surface. Kettle told WIRED that the model derived the hypothesis from proven examples, while he assessed and confirmed it. The model did not independently prove the new class.
That distinction is central to the event. The evidence does not show a system autonomously discovering and validating a broadly exploitable attack from start to finish. It shows that models can expand a skilled researcher's search space, surface connections and maintain parallel lines of inquiry when the human supplies domain boundaries and verification.
What security teams should take from it
For application-security teams, the practical lesson is to design AI-assisted research around provenance and falsification. Candidate findings need checks against prior work, reproducible evidence and human review before they become vulnerability claims. Security models can accelerate hypothesis generation, but confident-sounding novelty is not proof of novelty.
That is also the dual-use risk: the same collaboration pattern can help defenders find flaws and help attackers explore more paths. The Black Hat presentation and WIRED report support a near-term view of AI as a force multiplier for experts, with autonomous conceptual research still materially limited.
Key Points
- 1James Kettle presented a Black Hat study testing whether AI agents could produce novel web-security research.
- 2WIRED reports that autonomous performance remained limited, while structured human guidance generated many more useful research leads.
- 3The Shared-Parser Confusion hypothesis emerged from model analysis, but Kettle supplied the domain method and performed the validating work.
Scoring Rationale
The presentation offers direct evidence about both the capability and limits of AI-assisted vulnerability research, with immediate implications for offensive and defensive teams. The disclosed novel hypothesis is important, but broad exploitability and independent technical validation are not yet established.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
