Configure and launch an adversarial test campaign against your AI agent.
No agents registered. Register one first
Number of attack strategies per generation. Larger = more thorough but slower.
How many evolution cycles to run. More generations = deeper exploration.
Max back-and-forth exchanges per conversation. Longer = more complex attacks.
Choose which aspects of your agent to test. Each dimension uses specialized seed attacks and evaluation criteria.
Prompt injection, data leakage, jailbreaks
Hallucinations, consistency, reasoning
Policy, disclaimers, bias, content
The Attacker Model generates adversarial messages to probe your agent. The Judge Model evaluates conversations and scores attack effectiveness. Using different providers for each reduces shared blind spots.