Iggy
Your LLM serving stack is probably listening to the entire internet right now, and it will run whatever code that internet sends it. It will not ask who you are. Everyone is busy chasing prompt injection and jailbreaks. Meanwhile, the boring bugs are wide open. SGLang serves large language models for thousands of deployments, with 25k+ GitHub stars. I read the source and found two unauthenticated RCEs. There was nothing clever about it: a network-exposed broker hands attacker data straight to pickle.loads(). Send a payload, get a shell. I will walk the vulnerable code and pop a shell live on stage. You will leave knowing exactly what to grep for in your own AI stack, before someone else greps it for you. Two CVEs. One grep. Zero sophistication required.
Iggy
OX Security, Security Researcher
Igor Stepansky is a Security Researcher at OX Security, where he works on agentic AI for offensive security and automated penetration testing. Previously, he was part of Orca Security's Research Pod. His research spans unauthenticated remote code execution in AI/LLM serving frameworks, a use-after-free race condition in the Linux kernel's ksmbd SMB3 serve, and supply chain and CI/CD security across modern developer ecosystems. His work focuses on finding and exploiting high-impact vulnerabilities in widely deployed infrastructure, and he is a regular contributor of CVEs in open-source software.