Adversarial verification
Break AgentX
Try the requests the agent should never be allowed to execute. The point is not whether the model behaves — it is whether the control plane remains authoritative.
Attack the agent, not the customer
Adversarial dry-runs try to steal data, over-refund, duplicate actions, prompt-inject policy bypasses and remove Gemini. Nothing is executed.
Launch the adversarial suite to prove AgentX is governed even when the request is malicious or the model disappears.