An agent that can act can act wrongly. When it does, the consequence belongs to your organization, not to the model provider.
Capability does not establish readiness. Documentation does not establish readiness. Readiness is evidence: what the agent did under pressure, whether its context supports reliable behavior, whether your obligations became controls that hold, and whether you can authorize, watch, stop and reassess it.
You do not need a large program to start. You need one agent, one honest inventory, one adversarial scenario that tries to obtain something you have decided not to give away, and one written condition under which you would refuse to ship.
None of the incidents in Chapter 1 came from exotic model behavior. They came from authority nobody bounded, boundaries that existed only as assumptions, trust boundaries nobody defined, and evidence owned by the actor that produced it. Every one is fixable with ordinary engineering discipline, applied before the agent acts.
Before AI agents act, prove they are ready — not because proof makes failure impossible, but because it makes failure something your organization can see, explain, bound and answer for.