PortSwigger's HTTP Terminator: an autonomous agent that invented new desync techniques and used them on live banks.

James Kettle published the whitepaper August 5, presented at Black Hat and DEF CON (PortSwigger). The architecture detail is what agent builders should copy: Turbo Intruder got an MCP interface plus Python orchestration, and the exploitation agent's script was deliberately split so the LLM could not modify the deterministic segment that judges whether an attack succeeded. Actor separated from verifier, structurally. It brainstormed sixteen response-queue-poisoning hypotheses and autonomously evaluated each against every authorized bug-bounty target, yielding new desync triggers that compromised banks, security products, and government infrastructure. The answer to "can an agent invent, not just find" appears to be yes, when you refuse to let it grade its own work.