- James Kettle's AI turned 138 RFCs into roughly 30,000 attack ideas, and one payload worked on several unrelated web servers, including a bank's.
- Kettle said this finding worked because all the implementations are based off one RFC, and he said AI can do novel security research autonomously.
James Kettle, Director of Research at PortSwigger, said in his Black Hat USA 2026 talk that one payload can work on many products built from the same standard. His AI system mined standards at scale, and one of its finds showed this.
What was said
Kettle built an autonomous system called the HTTP Terminator. It read 138 RFCs, the public documents defining internet protocols, and produced roughly 30,000 new desync vectors. A desync attack breaks the separation between one web request and the next.
He tested them on about 30,000 bug bounty websites with one cloud server, and that step used no AI. It found proven desync flaws on maybe 700 sites.
The shared effect showed up in one finding. An RFC section covers how to answer a range request. The reply uses a multipart format. The Terminator proposed two places for a payload. One, after the closing marker in the body, worked on zero websites. The other, placing it straight in the body, worked on many sites, including unrelated web servers and a major bank Kettle could not name. Kettle explained that all the implementations are based off one RFC, so one payload can make many of them exploitable at the same time.
The airport and SSO cases differed. The Terminator also proposed gzip, another valid Transfer-Encoding value, paired with HTTP 1.0. It caused desync on many sites. His collaborator Paolo used it on an airport running F5 Big IP. A probe flagged odd behavior on a Meta server, and Paolo's manual scan then compromised systems including an SSO system.
Weaponizing attacks went badly at first. Kettle said models "replicate all the mistakes that beginners make, but they're harder to teach than beginners." He concluded: "AI can absolutely do novel security research autonomously."
Why it matters
Our reading: when many products implement the same standard, one payload can work on several of them at once, as Kettle described for the range-response finding. Buyers could ask which standards their products depend on.
The other side
The talk does not tie the airport or SSO findings to a shared RFC passage. Humans also did that exploitation.
The test pool was bug bounty sites, so the talk does not show how common these flaws are on the wider web. A bug in his own evaluation code also mangled valid payloads until he found it.
Kettle added that the system is a research tool and "it's not safe to run and you'd be absolutely mad to try and ship it in a commercial product".
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- Black Hat: Black Hat USA 2026 | Can AI Do Novel Security Research? Meet the HTTP Terminator (2026-11-18)





