run-adversarial-testing

Adversarial testing for AI agents involves executing automated multi-iteration sweeps that simulate bad-actor scenarios like prompt injection, PII extraction, and social engineering. This workflow validates agent robustness without requiring manual test set creation.

You should test for prompt injection, PII extraction, and social engineering vectors. Adversarial testing applies 10 distinct bad-actor scenarios including multi-turn escalation and compliance traps to verify the agent refuses unauthorized disclosures under emotional pressure.

Yes, you need the Coval CLI, an active API key, and a configured agent endpoint. These dependencies are required to coordinate test set creation, persona simulation, and composite metric evaluation during robustness sweeps.

Composite metric evaluation provides a per-scenario pass/fail scorecard that distinguishes between true security breaches and coherence issues. This intelligent scoring analyzes agent responses across simulated adversarial vectors to identify specific robustness failures.

Yes, adversarial testing applies to both voice and chat agents requiring validation against security vectors. The workflow coordinates persona simulation and composite metric evaluation to perform multi-iteration robustness sweeps across both agent types.