Research map / Adversarial robustness and security / Attacks on text and graphs
TAO-Attack: Toward Advanced Optimization-Based Jailbreak Attacks for Large Language Models
Zhi Xu, Jiaqi Li, Xiaotong Zhang, Hong Yu, Han Liu
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| replace I-GCG with TAO-Attack under fixed initialization | I-GCG · AdvBench, Llama-2-7B-Chat | ASR: 68% -> 92% | Improved |
| replace I-GCG with TAO-Attack under fixed initialization | I-GCG · AdvBench, Mistral-7B-Instruct-0.2 | iterations: 406 -> 86 | Improved |
| replace GCG update with DPTO (ablation) | GCG with harmful guidance · AdvBench 20 queries, Llama-2-7B-Chat | ASR: 55% -> 65% | Improved |
| replace GCG greedy update with softmax sampling (ablation) | GCG · AdvBench 20 queries, Llama-2-7B-Chat | ASR: 55% -> 55% | No clear effect |
| add Stage One refusal-aware loss and drop harmful guidance (ablation) | DPTO · AdvBench 20 queries, Llama-2-7B-Chat | ASR: 65% -> 100% | Improved |
| add Stage Two effectiveness-aware loss (ablation) | Stage One + DPTO · AdvBench 20 queries, Llama-2-7B-Chat | iterations: 261 -> 243 | Improved |
| use Rouge-L instead of embedding similarity for switching (ablation) | TAO-Attack · Llama-2-7B-Chat threshold 0.8 | ASR: 95% -> 100% | Improved |
| transfer universal suffix to GPT-3.5 Turbo | TAO-Attack vs GCG/I-GCG · Transfer | ASR: 30% -> 82% | Improved |
| use TAO-Attack on Qwen2.5-VL-7B | I-GCG · MM-SafetyBench | ASR: 90% -> 95% | Improved |
| TAO-Attack per-iteration overhead | I-GCG · per-iteration cost | time per iteration: 5.7s -> 7.1s (Stage Two) | Hurt |
Closest papers by meaning
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.93)
- Improved Techniques for Optimization-Based Jailbreaking on Large Language Models (ICLR 2025 · similarity 0.93)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.93)
- An Interpretable N-gram Perplexity Threat Model for Large Language Model Jailbreaks (ICML 2025 · similarity 0.92)
- Efficient LLM Jailbreak via Adaptive Dense-to-sparse Constrained Optimization (NeurIPS 2024 · similarity 0.91)
- Weak-to-Strong Jailbreaking on Large Language Models (ICML 2025 · similarity 0.91)
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.
