Promp2Pwn – LLMs Winning at Pwn2Own
Georgi G
[un]prompted 2026 — AI Security Practitioner Conference · Day 2 · 1
AI review
A Pwn2Own win using an LLM agent to find the bugs is not a thought experiment — it happened, Geshev has the logs, the CVEs, and the demo. The technical depth here is real: JADX MCP integration, LangChain + LiteLLM + LangFuse architecture, inline de-obfuscation agent, multi-run aggregation, and a bug verifier that almost worked against him.