# Claim 4: Static-Attacker Ablation Study --- We verified Claim 4 from Section 5.3 (Figure 3): - **Safety gain**: Compared to training against a static attacker (where jailbreaks are fixed), co-evolution training against a dynamic attacker yields a **12% reduction** in Attack Success Rate (ASR) under the NetSafe threat. - **Utility gain**: AdvEvo-MARL also achieves up to a **4% improvement** in task performance (specifically on LiveCodeBench) compared to the static attacker baseline. - **Underlying reason**: Evolving attackers continuously discover new, diverse jailbreaks (non-monotonic increasing trend in attacker prompt diversity shown in Figure 4). This acts as a regularizer, preventing defenders from overfitting to specific jailbreak patterns and encouraging the emergence of generalizable task-solving and defense capabilities.