Research map / Adversarial robustness and security / Attacks on text and graphs
Jailbreaking the Matrix: Nullspace Steering for Controlled Model Subversion
Vishal Pramanik, Maisha Maliha, Susmit Jha, Sumit Kumar Jha
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 |
|---|---|---|---|
| add head masking to nullspace steering (ablation) | HMNS · AdvBench, Phi-3 Medium 14B | ASR (%): 89.5 -> 96.8 | Improved |
| add nullspace projection to head masking (ablation) | HMNS · AdvBench, Phi-3 Medium 14B | ASR (%): 87.9 -> 96.8 | Improved |
| inject orthogonal nullspace direction instead of direct direction (ablation) | HMNS · AdvBench, Phi-3 Medium 14B | ASR (%): 88.7 -> 96.8 | Improved |
| re-identify top-K heads at each decoding attempt (ablation) | HMNS · AdvBench, Phi-3 Medium 14B | ASR (%): 90.2 -> 96.8 | Improved |
| select heads by KL attribution instead of random (ablation) | HMNS · AdvBench, Phi-3 Medium 14B | ASR (%): 81.4 -> 96.8 | Improved |
| intervene on multiple layers instead of one (ablation) | HMNS · AdvBench, Phi-3 Medium 14B | ASR (%): 86.1 -> 96.8 | Improved |
| inject at multiple positions instead of final token only (ablation) | HMNS · AdvBench, Phi-3 Medium 14B | ASR (%): 96.8 -> 95.0 | Mixed |
| use HMNS vs best baseline ArrAttack | HMNS · AdvBench LLaMA-2-7B-Chat | ASR (%): 92.00 -> 98.00 | Improved |
| use HMNS vs ArrAttack under six defenses | HMNS · AdvBench LLaMA-2-7B-Chat defenses average | ASR (%): 49.3 -> 55.8 | Improved |
| use HMNS vs ArrAttack compute | HMNS · AdvBench LLaMA-3.1-70B | FLOPs per success (x10^12): 0.62 -> 0.53 | Improved |
| partial masking instead of hard masking (ablation) | HMNS · Phi-3 Medium 14B AdvBench | ASR | Hurt |
| KL attribution instead of simpler heuristics (logit drop, entropy) (ablation) | HMNS · Phi-3 Medium 14B AdvBench | ASR: lose 5-8 points of ASR | Improved |
Closest papers by meaning
- Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis (NeurIPS 2026 · similarity 0.89)
- Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks (CVPR 2025 · similarity 0.88)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.88)
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.87)
- Steering Externalities: Benign Activation Steering Unintentionally Increases Jailbreak Risk for Large Language Models (NeurIPS 2026 · similarity 0.87)
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense (NeurIPS 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.
