AI Coding Agent Horror Stories: Security Risks Explained

A detailed examination of six documented security failure categories in AI coding agents, illustrated with real incidents from 2024-2026. Covers filesystem wipes (Claude Code deleting home directories), excessive privilege inheritance (AWS Kiro taking down Cost Explorer for 13 hours), secrets leakage via agent context (the s1ngularity npm supply chain attack weaponizing AI agents for reconnaissance), prompt injection leading to private key exfiltration, malicious plugin supply chains (ClawHavoc campaign distributing Atomic Stealer via 335+ fake OpenClaw skills), and autonomous irreversible actions (Replit agent wiping a production database during a code freeze). For each category, the post explains why it happens structurally and how Docker Sandboxes mitigates it through microVM isolation, blocked credential paths, network egress controls, and Git worktree-scoped execution.