Research map / Adversarial robustness and security / Attacks on text and graphs
Endless Jailbreaks with Bijection Learning
Brian R.Y. Huang, Maximilian Li, Leonard Tang
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 |
|---|---|---|---|
| Apply bijection learning instead of the strongest listed baseline | Frontier models on AdvBench-50 · AdvBench-50 | ASR | Improved |
| Apply bijection learning instead of the strongest listed baseline | Frontier models on HarmBench · HarmBench test set | ASR | Improved |
| Increase the bijection attack budget | Bijection learning · AdvBench and HarmBench | ASR | Improved |
| Increase dispersion from low to a middle range (ablation) | Bijection learning · HarmBench-35 | ASR | Improved |
| Increase dispersion beyond the effective middle range (ablation) | Bijection learning · HarmBench-35 | ASR | Hurt |
| Use letter-to-digit rather than letter-to-letter bijections (ablation) | Bijection learning on Claude 3 Haiku and GPT-4o-mini · HarmBench-35 | ASR | Hurt |
| Increase dispersion and encoding length (ablation) | Bijection learning · 10-shot MMLU | MMLU score | Hurt |
| Use the selected single bijection rather than an average-case bijection | Bijection learning · HarmBench | ASR | Improved |
| Human-filter responses marked unsafe by the LLM judge | Bijection learning evaluation · HarmBench | ASR | Mixed |
| Use GPT-4o-mini input and output guards with Claude 3.5 Sonnet | Guarded Claude 3.5 Sonnet · HarmBench | ASR: reduces ASR by 9 percentage points | Hurt |
| Use Claude 3 Haiku input and output guards with Claude 3.5 Sonnet | Guarded Claude 3.5 Sonnet · HarmBench | ASR: near-zero | Hurt |
| Use Claude 3.5 Sonnet input and output guards with Claude 3.5 Sonnet | Guarded Claude 3.5 Sonnet · HarmBench | ASR: reduces ASR by 10 percentage points | Hurt |
Closest papers by meaning
- Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space (NeurIPS 2024 · similarity 0.89)
- Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation (ICLR 2024 · similarity 0.89)
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense (NeurIPS 2025 · similarity 0.89)
- Assessing Vulnerabilities in State-of-the-Art Large Language Models Through Hex Injection (Student Abstract) (AAAI 2025 · similarity 0.89)
- One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs (ICLR 2025 · similarity 0.89)
- Understanding and Enhancing the Transferability of Jailbreaking Attacks (ICLR 2025 · similarity 0.89)
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.
