ReaLM - PromptInject: leaderboard

Metric: Attack success rate (%; share of adversarial cases where the model gives an incorrect, task-relevant answer) under PromptInject (adversarial instructions appended to the prompt) on the 832 ReaLM physical-world cases (driving, manipulation, grasping, physics, egocentric, diagnostics and scene QA; image, question and ground-truth answer), black-box one-shot evaluation at temperature 0, a Qwen3-8B judge extracting the answer; lower is better. Source: arxiv.org. Saturation forecast: Around 2033. 13 models tracked.

Top models

#ModelScore
1Seed 2.0 Mini52.9
2GPT-4.1 Mini55.4
3Kimi K2.559.5
4Gemini 3 Flash61.9
5Qwen 3.5 9B63.7
6Qwen 3.6 Flash64
7Qwen 3 VL 32B64
8Qwen 3.5 122B A10B64.3
9Qwen 3.5 27B66

Interactive version: theaggregate.ai/benchmark?slug=realm-promptinject · How It Works · Data refreshed daily, snapshot 2026-09-29.