When Agents Execute: RCE Paths in LLM-Powered Coding Tools

This talk is a follow-on to our September 2025 research on denial-of-service and permission escape in Claude Code. We now examine how LLM-powered coding agents can be weaponized end-to-end, including paths to remote code execution. Using Claude Code as a primary case study, and extending to VS Code extension exploits and recent Cursor incidents, we show how agent autonomy, extension APIs, and execution boundaries collapse into a practical RCE surface.