AI Paper Atlas

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

ICML 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: LLM jailbreak and injection defense

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 changedWhereMeasuredEffect
add honeypot preference trainingThreshold defense, Llama-3 8B · Targeted embedding attacksensemble ASR: 51.6 -> 30.6Improved
add honeypot preference trainingThreshold defense, Qwen-3 8B · Targeted embedding attacksensemble ASR: 33.1 -> 25.9Improved
add honeypot preference trainingCircuit Breaking, Llama-3 8B · Targeted embedding attacksensemble ASR: 51.8 -> 46.5Improved
add honeypot preference trainingCircuit Breaking, Qwen-3 8B · Targeted embedding attacksensemble ASR: 45.8 -> 32.6Improved
add honeypot preference trainingCircuit Breaking, Llama-3 8B · GRPO adaptive attackerensemble ASR: 47.4 -> 54.6Hurt
add honeypot preference trainingThreshold defense, Llama-3 8B · GRPO adaptive attackerensemble ASR: 78.6 -> 59.9Improved
add honeypot preference trainingCircuit Breaking, Qwen-3 8B · GRPO adaptive attackerensemble ASR: 72.8 -> 55.1Improved
add honeypot preference trainingThreshold defense, Llama-3 8B · Harmful prefix-tuning attackensemble ASR: 80.0 -> 87.0Hurt
add honeypot preference trainingCircuit Breaking, Llama-3 8B · Harmful prefix-tuning attackensemble ASR: 77.0 -> 92.0Hurt
add honeypot preference trainingCircuit Breaking, Qwen-3 8B · GRPO jailbreak outputs, human ratinghuman actionability (0-4): 1.95 -> 0.84Improved
add honeypot preference trainingThreshold defense, Qwen-3 14B · Soft-prompt embedding attacksensemble ASR: 19.2% to 8.1%Improved
add DPO honeypot preference (versus reversed preference) (ablation)Threshold defense · Targeted embedding attacksconditional actionability: 59.67 -> 46.53Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy