REPRO: Black-Box Evolutionary Prompt Optimization Enables LLM Jailbreaks
Start here
An arXiv paper released on October 11, 2026 proposes REPRO, a pure black-box automated red-teaming framework that uses evolutionary optimization driven by a compressed open-source 7B reflection model on a single consumer GPU to search for jailbreak prompts. The method introduces two mechanisms: LLM-guided multi-parent recombination (semantic crossover), and Bayesian policy selection via Thompson Sampling informed by structured failure analysis. The authors report coverage of 36 target models across 11 model families on AdvBench, JailbreakBench, and HarmBench, including GPT-5-nano, Clau (original text truncated here).
Generated by AI from reports · Updated 29 days ago
Attention and developments
Attention, hourly
Now 34, 0% in 24 hours·Peak 34 (Nov 6 03:00)
Related research
This event's paper
- REPRO: Reflective Evolutionary Prompt Optimizer for Automated LLM Red-Teaming
arXiv · 2026-10-11 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningREPRO: Reflective Evolutionary Prompt Optimizer for Automated LLM Red Teaming
The paper proposes REPRO, a purely black-box automated red-teaming framework that drives evolutionary optimization with a compressed open-source 7B reflection model on a single consumer GPU to search for jailbreak prompts. The method introduces two mechanisms: LLM-guided multi-parent recombination, i.e., semantic crossover, and Bayesian policy selection via Thompson Sampling based on structured failure analysis. The authors report covering 36 target models across 11 model families on AdvBench, JailbreakBench, and HarmBench, including GPT-5-nano, Claude-3.5-Sonnet, Gemini-2.5-Pro, and Grok-4.1, achieving 100% attack success rate on each evaluation target in a single round, with a median of 24 target queries per behavior.