Safety reports for GPT-5.2, Gemini 3 Pro, Qwen3-VL, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5

The researchers conducted a safety evaluation under a unified protocol on six frontier models—GPT-5.2, Gemini 3 Pro, Qwen3-VL, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5—covering language, vision-language, and image generation, combining benchmark, adversarial, multilingual, and compliance tests, and summarized the results into a safety leaderboard and model profiles. The results show that safety levels are highly uneven: GPT-5.2 performs consistently strongly and evenly, while other models show clear trade-offs among benchmark safety, adversarial robustness, multilingual generalization, and compliance. Despite good scores on standard benchmarks, all models remain highly vulnerable under adversarial testing, with the worst-case safety rate below 6%. Text-to-image models align somewhat more strongly on regulated visual risk categories but remain vulnerable to adversarial or semantically ambiguous prompts. The report emphasizes that frontier model safety is influenced by multiple dimensions such as modality, language, and evaluation design, and that standardized, holistic safety evaluation is needed to reflect real risks.