Autonomous coding assistants like Cursor, Windsurf, and Claude Code have evolved from passive autocomplete into agentic executors. But when these agents operate within local development environments, they inherit a terrifying identity scope: access to SSH keys, cloud tokens, Supabase/AWS credentials, and CI/CD pipelines.

This session maps the attack paths introduced by agentic AI in the developer loop. Drawing from recent experience orchestrating a complex React Native and Supabase architectural overhaul using multiple agentic tools, we will demonstrate how a hallucinating, poisoned, or compromised AI assistant can traverse a developer's local identity permissions to escalate privileges into production cloud infrastructure.

We will model the AI Dev Loop as a graph, mapping the edges between local agent execution and enterprise identity compromise. Attendees will learn how to threat-model these hybrid attack paths, define strict trust boundaries for autonomous dev tools, and implement detection workflows to identify anomalous identity usage originating from agentic workflows.