Target the actual risk
Adversarial tests should be derived from the agent contract. A sales agent should be tested for invented pricing and guarantees; a support agent should be tested for sensitive-data handling and unsupported state changes.
Include indirect requests
Do not only ask an agent to violate policy directly. Test requests disguised as troubleshooting, urgency, authority, or a request to ignore previous instructions.
Treat findings as engineering input
A useful failed test includes the scenario, the expected behavior, the actual response, severity, and an explanation. The result should be actionable for the agent team, not merely a score reduction.