Coding Agent Horror Stories: The 29 Million Secret Problem
SMRTR summary
A poisoned npm package (s1ngularity) hijacked installed AI coding agents like Claude Code and Gemini CLI by invoking them with permission-bypass flags to scan and exfiltrate credentials — no exploit needed. Docker Sandboxes addresses this by isolating agents in microVMs with workspace-scoped filesystems and proxy-injected secrets, ensuring credentials never enter the agent's reach.
SMRTR provides this summary for quick context. The original article belongs to Daily.dev.
Read the original article