Research map / Adversarial robustness and security / Attacks on text and graphs
Training with Honeypots: Reshaping How LLMs Fail Under Adversarial Attacks
Samuel Simko, Punya Syon Pandey, Zhijing Jin, Bernhard Schölkopf
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 honeypot preference training | Threshold defense, Llama-3 8B · Targeted embedding attacks | ensemble ASR: 51.6 -> 30.6 | Improved |
| add honeypot preference training | Threshold defense, Qwen-3 8B · Targeted embedding attacks | ensemble ASR: 33.1 -> 25.9 | Improved |
| add honeypot preference training | Circuit Breaking, Llama-3 8B · Targeted embedding attacks | ensemble ASR: 51.8 -> 46.5 | Improved |
| add honeypot preference training | Circuit Breaking, Qwen-3 8B · Targeted embedding attacks | ensemble ASR: 45.8 -> 32.6 | Improved |
| add honeypot preference training | Circuit Breaking, Llama-3 8B · GRPO adaptive attacker | ensemble ASR: 47.4 -> 54.6 | Hurt |
| add honeypot preference training | Threshold defense, Llama-3 8B · GRPO adaptive attacker | ensemble ASR: 78.6 -> 59.9 | Improved |
| add honeypot preference training | Circuit Breaking, Qwen-3 8B · GRPO adaptive attacker | ensemble ASR: 72.8 -> 55.1 | Improved |
| add honeypot preference training | Threshold defense, Llama-3 8B · Harmful prefix-tuning attack | ensemble ASR: 80.0 -> 87.0 | Hurt |
| add honeypot preference training | Circuit Breaking, Llama-3 8B · Harmful prefix-tuning attack | ensemble ASR: 77.0 -> 92.0 | Hurt |
| add honeypot preference training | Circuit Breaking, Qwen-3 8B · GRPO jailbreak outputs, human rating | human actionability (0-4): 1.95 -> 0.84 | Improved |
| add honeypot preference training | Threshold defense, Qwen-3 14B · Soft-prompt embedding attacks | ensemble ASR: 19.2% to 8.1% | Improved |
| add DPO honeypot preference (versus reversed preference) (ablation) | Threshold defense · Targeted embedding attacks | conditional actionability: 59.67 -> 46.53 | Improved |
Closest papers by meaning
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal (ICML 2024 · similarity 0.90)
- Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.90)
- WildTeaming at Scale: From In-the-Wild Jailbreaks to (Adversarially) Safer Language Models (NeurIPS 2024 · similarity 0.89)
- Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge (NeurIPS 2026 · similarity 0.89)
- Expected Harm: Rethinking Safety Evaluation of (Mis)Aligned LLMs (NeurIPS 2026 · similarity 0.89)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.89)
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.
