Why Prompt-Injection Boundaries Fail Once Your Agent Can Run Shell Commands

Prompt injection against a shell-capable agent is a process problem: model text becomes argv. This short shows the vulnerable exec wrapper, a terminal run where an appended command executes, and the executor controls that hold: argument arrays, binary and flag allowlists, fixed cwd, containers, no secrets or network, and treating stdout as untrusted. Full write-up: https://hackyjs.com/posts/why-prompt-injection-boundaries-fail-once-your-agent-can-run-shell-commands

Read the post →