Agent Security Bench (ASB) security claim for the ASB target: one task per (agent, benign task, attacker tool), with ASB's attack-success / utility / refusal predicates ported verbatim and hierarchical factories by scenario, aggressiveness, and attack type.