Attackers can inject indirect prompts in normal-looking repositories to trick Claude Code into spawning a reverse shell.

One condition did what my nagging couldn't ...

All my agents needed was a little bit of codified workflows to follow ...