Summary

Tested curated top-ASR prompts against two frontier-scale models: NVIDIA Nemotron 3 Super (~230B) and Alibaba Qwen3.5 (397B). Parameter scale alone does not predict safety robustness. Nemotron 230B showed 78.6% heuristic ASR while Qwen3.5 397B showed 7.1% corrected ASR, including a novel ‘silent refusal’ pattern.