In autonomous agent security, in-band guardrails fail because reasoning models granted local execution permissions treat…