Variable ATTACK_CATEGORIESConst
ATTACK_CATEGORIES: readonly [
"prompt-injection",
"jailbreak",
"data-exfiltration",
"privilege-escalation",
"output-manipulation",
"context-poisoning",
"resource-exhaustion",
"social-engineering",
"multi-step-attack",
"cross-agent-exploitation",
] = ...
Adversarial Sample Active Generation — R4-04
Proactively generates adversarial test cases based on agent profile, known weaknesses, and attack taxonomy.