Research map / Adversarial robustness and security / Attacks on text and graphs
New Wide-Net-Casting Jailbreak Attacks Risk Large Models
Qiuchi Xiang, Haoxuan Qu, Hossein Rahmani, Jun Liu
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| wide-net-casting (OR over 4 LLMs) instead of single model | GCG · AdvBench | WASR vs ASR: 46.2% -> 75.0% | Improved |
| wide-net-casting instead of single model | ReMiss · AdvBench | WASR vs ASR: 86.5% -> 92.3% | Improved |
| wide-net-casting over MLLMs | MLAI+PixArt-alpha · MM-SafetyBench | WASR vs ASR: 77.9% -> 93.7% | Improved |
| add SmoothLLM defense | ReMiss group of 4 LLMs · AdvBench | WASR: 92.3% -> 61.5% | Hurt |
| add RobustKV defense | ReMiss group of 4 LLMs · AdvBench | WASR: 92.3% -> 56.1% | Hurt |
| add VLGuard defense | MLAI+PixArt-alpha group · MM-SafetyBench | WASR: 93.7% -> 40.2% | Hurt |
| restrict group to same LLaVA family | MLAI+PixArt-alpha · MM-SafetyBench | WASR: 93.7% -> 89.9% | Hurt |
| restrict group to same family and version | MLAI+PixArt-alpha · MM-SafetyBench | WASR: 89.9% -> 41.1% | Hurt |
| add joint specialized training (proposed) | ReMiss baseline · AdvBench LLMs | WASR: 92.3% -> 100% | Improved |
| add joint specialized training (proposed) | MLAI+PixArt-alpha · AdvBench MLLMs + VLGuard | WASR: 37.5% -> 50.8% | Improved |
| add theoretically derived update schedule over rule-based variants (ablation) | Variant III · AdvBench MLLMs | WASR: 96.7% -> 100% | Improved |
| add linearly decreasing H_t instead of fixed H_t (ablation) | Variant V · AdvBench MLLMs | WASR: 98.0% -> 100% | Improved |
Closest papers by meaning
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.89)
- h4rm3l: A Language for Composable Jailbreak Attack Synthesis (ICLR 2025 · similarity 0.88)
- Bag of Tricks: Benchmarking of Jailbreak Attacks on LLMs (NeurIPS 2024 · similarity 0.88)
- One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs (ICLR 2025 · similarity 0.88)
- Jailbreak susceptibility prediction and mitigation via the behavioral geometry of models (NeurIPS 2026 · similarity 0.87)
- An Interpretable N-gram Perplexity Threat Model for Large Language Model Jailbreaks (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
