“Please Don't Delete My Repo”: Why You Can't Trust Your AI Coding Agent

Abstract

Telling your AI coding agent “please don't commit directly to main” is about as effective as asking your cat not to knock things off the table. LLMs are non-deterministic by nature — a polite prompt instruction is not a security boundary. In this session, you'll learn how to enforce what you can't simply ask for: lifecycle hooks that intercept tool calls before they execute, blocking unauthorized Git commits, file deletions, or rogue shell commands at the architectural level. We'll also explore how to run coding agents autonomously and safely inside sandboxes.