Containment of a prompt-injected agent
An agent that follows injected instructions cannot weaken verification — a rewritten verification config is ignored because configuration is read from the base commit and the path is protected — and a host secret behind a planted symlink never reaches a context pack.