We’ve just released our May signature prompt pack with over 10,500 new adversarial prompts covering 58 harmful intents. This prompt pack includes a new attack vector: Scenario nesting.

Scenario nesting is a jailbreak strategy where a harmful or policy-violating instruction is embedded inside a familiar and benign-looking task prompt—such as code completion, table generation, or paragraph continuation—in order to bypass safety filters.

Scenario nesting can also be used to generate agent attack prompts for Agentic Warfare campaigns.