Research map / Adversarial robustness and security / Backdoor and poisoning
Rapid Poison: Practical Poisoning Attacks Against the Rapid Response Framework
David Huang, Jaewon Chang, Avidan Shah, Prateek Mittal, Chawin Sitawarin
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 few-shot MCQ examples to the injection instead of zero-shot description (ablation) | Format poisoning · MMLU domains | FPR: 94-100% FPR, ~30-60% above zero-shot | Improved |
| use zero-shot format description up to 5% poisoning (ablation) | Format poisoning · MMLU domains | FPR: many domains below 60-70% | Hurt |
| use held-out WildChat queries as poison instead of unrelated synthetic safe queries (ablation) | ChatGPT name poisoning · Name-based FPR on Claude | FPR: 48% -> 19% | Improved |
| use contrastive query augmentation instead of naive few-shot nouns (ablation) | ChatGPT name poisoning · Name-based FPR on Claude | FPR: 70% -> 19% | Improved |
| add LLM-guided concept embedding instead of naive HP name insertion (ablation) | Omission attack HP · PAIR jailbreaks | FNR: 14.7% -> 55.3% | Improved |
| use paraphrased longer trigger instead of short n-gram trigger (ablation) | Omission attack human-like style · Cipher jailbreak | FNR: 0% -> 40.67% | Improved |
| use paraphrased trigger Meshy instead of Midjourney (ablation) | Omission attack · Harmful queries | FNR: 96% -> 98% | No clear effect |
| add greedy 3-gram set-cover sampling to injection (ablation) | Distribution poisoning · low poisoning rates | FPR | Improved |
| use Meta SecAlign 70B as proliferation model (ablation) | Domain poisoning · Professional law FPR | FPR: 98% -> 0% | Improved |
| raise poisoning rate from 1% in omission (ablation) | Direct poisoning vs RR pipeline · Harmful queries | FNR: 62-63% harmful vs 94-100% jailbreaks | Mixed |
| poison SWE-Bench coding prompts then test on WildChat | Distribution poisoning · WildChat | FPR: 100% -> 0% | No clear effect |
Closest papers by meaning
- h4rm3l: A Language for Composable Jailbreak Attack Synthesis (ICLR 2025 · similarity 0.88)
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.87)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.86)
- Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks (ICLR 2025 · similarity 0.86)
- Understanding and Enhancing the Transferability of Jailbreaking Attacks (ICLR 2025 · similarity 0.85)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.85)
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.
