Research map / Large language models / Language, safety and interpretability
PLAGUE: Plug-and-play Framework for Lifelong Adaptive Generation of Multi-turn Exploits
Neeladri Bhuiya, Madhav Aggarwal, Diptanshu Purwar
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 backtracking (ablation) | GOAT finisher · HarmBench, target OpenAI o3 | SRE@2: 0.587 -> 0.612 | Improved |
| add reflection (ablation) | GOAT + BT · HarmBench, target OpenAI o3 | SRE@2: 0.612 -> 0.761 | Improved |
| add planner (ablation) | GOAT + BT + R · HarmBench, target OpenAI o3 | Bin-ASR@2: 0.59 -> 0.582 | No clear effect |
| add retrieval of successful strategies (ablation) | GOAT + BT + R + P · HarmBench, target OpenAI o3 | SRE@2: 0.773 -> 0.814 | Improved |
| add backtracking (ablation) | GOAT finisher · HarmBench, target Claude Opus 4.1 | SRE@2: 0.222 -> 0.396 | Improved |
| add reflection (ablation) | GOAT + BT · HarmBench, target Claude Opus 4.1 | SRE@2: 0.396 -> 0.402 | Improved |
| add planner (ablation) | GOAT + BT + R · HarmBench, target Claude Opus 4.1 | SRE@2: 0.402 -> 0.431 | Improved |
| add retrieval of successful strategies (ablation) | GOAT + BT + R + P · HarmBench, target Claude Opus 4.1 | SRE@2: 0.431 -> 0.465 | Improved |
| replace ActorBreaker planner with PLAGUE planner (ablation) | Crescendo + BT + R · HarmBench, target Claude Opus 4.1 | SRE@2: 0.601 -> 0.673 | Improved |
| add BT, reflection and ActorBreaker planner to Crescendo (ablation) | Crescendo · HarmBench, target Claude Opus 4.1 | SRE@2: 0.48 -> 0.601 | Improved |
| use Crescendo instead of GOAT as finisher (ablation) | PLAGUE · HarmBench, target Claude Opus 4.1 | SRE@2: 0.465 -> 0.673 | Improved |
| PLAGUE vs best baseline GOAT | GOAT · HarmBench, target OpenAI o3 | SRE@2: 0.587 -> 0.814 | Improved |
Closest papers by meaning
- MT-JailBench: A Modular Benchmark for Understanding Multi-Turn Jailbreak Attacks (NeurIPS 2026 · similarity 0.91)
- Multi-Turn Jailbreaking Large Language Models via Attention Shifting (AAAI 2025 · similarity 0.91)
- MetaCipher: A Time-Persistent and Universal Multi-Agent Framework for Cipher-Based Jailbreak Attacks for LLMs (AAAI 2026 · similarity 0.90)
- WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models (NeurIPS 2024 · similarity 0.89)
- Helpful to a Fault: Measuring Illicit Assistance in Multi-Turn, Multilingual LLM Agents (ICML 2026 · similarity 0.89)
- Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation (ICLR 2024 · similarity 0.88)
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.
