ResearchBenchmarksPrompt Injection Attack on Office Real-world (test)Follow54.2ASRPI3D53.78456.59259.462.208Feb 6, 2026Evaluation ResultsMethodMethodLinksASRPI3DEvaluator MLLM=GPT-4o-...Evaluator MLLM=GPT-4o-mini2026.0254.2PI3DEvaluator MLLM=Gemini-...Evaluator MLLM=Gemini-2.0-Flash2026.0264.6