RoboJailBench: leaderboard

Harmonic mean (%) of the security rate (adversarial goals rejected) and the utility rate (benign goals accepted) in RoboJailBench (2026), where each robot-scene image from six embodied datasets comes with a benign and an adversarial instruction and the vision-language model must answer okay or deny. Higher is better.

Metric: Security-utility harmonic mean (%) of the rejection rate on adversarial goals and the acceptance rate on benign goals, over the paired benign and adversarial goals of six robot-scene datasets (RoboVQA, DROID, NVIDIA PhysicalAI AV, RH20T and Robo2VLM augmented with intent-contrast prompts, and the new RJB-Instructions), each image with one benign and one adversarial instruction and a forced okay or deny answer, no attack and no defense; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.4 Nano (2026-03-17)97.84#473
2Claude Haiku 4.596.57#583
3GPT-5.4 Mini95.93#406
4Gemini 3 Flash (Preview)Non-reasoning95.37#152
5gemini-robotics-er-1.6-previewNon-reasoning92.75

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=robojailbench · How It Works · Data refreshed daily, snapshot 2026-11-21.