Research map / Large language models / Language, safety and interpretability
SpatialJB: How Text Distribution Art Becomes The "Jailbreak Key" for LLM Guardrails
Zhiyi Mou, Jingyuan Yang, ZEHENG QIAN, Wangze Ni, Tianfang Xiao, Ning Liu, Chen Jason Zhang, Zhan Qin, Kui Ren
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 |
|---|---|---|---|
| reformat harmful output into 2D spatial layouts (SpatialJB) instead of Base64, Zulu or Ubbi Dubbi | GPT-4 · Repeat jailbreak, no guardrail | ASR (%): 7.7%-12.6% -> 95%-98.3% | Improved |
| use SpatialJB instead of SATA | GPT-4 · Repeat jailbreak, no guardrail | ASR (%): 25.8% -> 95% (Acrostic) | Improved |
| add Llama Guard 4 output guardrail | SpatialJB on GPT-4 · Repeat jailbreak | ASR (%): 95% -> 91% | Hurt |
| add OpenAI Moderation output guardrail | SpatialJB on Gemini 2.5-Pro · Repeat jailbreak | ASR (%): 93.7% -> 77% | Hurt |
| switch from Repeat task to Summary task (ablation) | SpatialJB · Repeat vs Summary jailbreak | ASR (%): 97.3% -> 73% (GPT-4 Staircase) | Hurt |
| target Claude instead of other LLMs (ablation) | SpatialJB · Repeat jailbreak, no guardrail | ASR (%): 100% (Deepseek-V3) -> 68% (Claude Acrostic) | Hurt |
| combine SpatialJB with Positive Guidance and Think Guidance (ablation) | SpatialJB on Claude · Repeat jailbreak ablation | ASR (%): 68 -> 97.60 | Improved |
| apply Positive Guidance or Think Guidance alone without format control (ablation) | Claude · Repeat jailbreak ablation | ASR (%): 68 -> 20.80 and 10.40 (single modules) | Hurt |
| apply SpatialJB on top of FlipAttack outputs | FlipAttack · AdvBench, Llama Guard | harmful outputs reclassified as safe: 415 of 521 (79.65%) became safe-classified | Improved |
| spatial 2D layout instead of linear layout | token semantic similarity · embedding similarity example | semantic correlation: 0.9419 -> 0.5846 | Hurt |
Closest papers by meaning
- Position: Building Guardrails for Large Language Models Requires Systematic Design (ICML 2024 · similarity 0.87)
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.87)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.87)
- Jailbreaking Large Language Models Against Moderation Guardrails via Cipher Characters (NeurIPS 2024 · similarity 0.87)
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense (NeurIPS 2025 · similarity 0.87)
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 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.
