Research map / Adversarial robustness and security / Black-box and transfer attacks
Jailbreak Transferability Emerges from Shared Representations
Rico Angell, Jannik Brinkmann, He He
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 |
|---|---|---|---|
| add benign-only distillation from Gemma2-27B | Qwen2.5-14B · Gemma2-27B teacher similarity | mutual kNN similarity: 0.28 -> 0.42 | Improved |
| add benign-only cross-family distillation | Student models · Transfer of StrongREJECT jailbreaks to teacher | mean transfer score | Improved |
| add benign-only distillation then GCG attack | Distilled students · Adaptive GCG jailbreaks | transferability | Improved |
| add in-family distillation (Llama3.1-8B into Llama3.2-3B) | Llama3.2-3B · Model similarity | mutual kNN similarity | No clear effect |
| add benign-only distillation from Llama3.1-8B | Qwen2.5-7B · AIM persona jailbreak | jailbreak strength | Improved |
| use persona-style jailbreaks instead of direct ones | Jailbreak transfer · Strength >= 0.3 | models reached: 8% -> 47% | Improved |
| use persona-style jailbreaks instead of direct ones | Jailbreak transfer · Strength >= 0.5 | models reached: 4% -> 17% | Improved |
| use higher source-target representational similarity | Model pairs · 380 model pairs | symmetric transfer AUROC | Improved |
Closest papers by meaning
- Understanding and Enhancing the Transferability of Jailbreaking Attacks (ICLR 2025 · similarity 0.90)
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.88)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.87)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.87)
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense (NeurIPS 2025 · similarity 0.87)
- AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion (NeurIPS 2026 · 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.
