您的当前位置：首页 > 行业数据 > harmful outputs (FMF, 2025).

harmful outputs (FMF, 2025).

行业数据

下载图片

原图定位

同报告图片

/ 4

所属报告：

GovAI：2026构建前沿人工智能生物滥用防护措施统一评估标准技术报告（英文版）（35页）.pdf

打包全文图表

safeguard success; it must be weighed against the false-positive rate.6 SecureBio's recently developed BioTIER benchmark is a step in the right direction, providing prompts categorized across three biological risk sets? (SecureBio, 2026). The benchmark prompts that should be refused (BioTIER-refuse). Within the refused category, a subset of prompts is additionally tagged as Select Agent content, meaning they involve biological agents or toxins subject to heightened regulatory concern. Applying BioTIER to individual frontier 6 Beyond the false-positive rate of prompt handling, a more detailed benchmark or evaluation could consider other costs such as latency or reduced access for legitimate scientific researchers. 7 These are: Catastrophe Avoidance (CA), Biomedical Dual-Use Research of Concern (BD), and Related Biology (RB). CA and BD contain prompts that models should refuse, while RB contains benign biology prompts that models should answer