OpenAI发布了GPT-Rosalind-6的系统卡附录 1。该模型是基于GPT-6 Astra权重开发的前沿推理模型,专门用于生物学、药物发现和转化医学 1。
Rosalind采用“受信任访问”模式,仅通过受控访问向经过审查的合格科学家和机构提供 1。申请人在获得权限前,需接受对其拟议用途及安全控制措施的审核 1。
在能力框架下,根据Preparedness Framework评估,该模型在生物化学领域被评为高能力(High capability),低于关键(Critical)阈值 1。然而,该模型在网络安全领域被视为关键风险(Critical),而在AI自我改进领域则低于高能力水平 1。尽管Rosalind对双重用途生物查询的拒绝率较低,但模型仍保留对生物武器化请求的拒绝功能,并在其他安全主题上维持常规的拒绝策略 1。
OpenAI has published a system card addendum for GPT-Rosalind-6, an advanced reasoning model developed using the same weights as the flagship GPT-6 Astra 1. The model is specialized for applications in biology, drug discovery, and translational medicine 1. Access to GPT-Rosalind-6 is restricted through a "trusted access" framework, which grants permissions only to vetted qualified scientists and institutions after reviewing their proposed use cases and security controls 1.
Under the Preparedness Framework, the model’s capabilities are assessed across multiple domains. In biochemistry, GPT-Rosalind-6 is rated as having High capability, which remains below the Critical threshold 1. In the domain of AI self-improvement, its performance is evaluated as being below the High capability level 1. However, the model is considered to pose a Critical risk in cybersecurity 1. While GPT-Rosalind-6 demonstrates a lower refusal rate for dual-use biological queries compared to general models, it maintains strict refusals for requests related to biological weaponization and adheres to standard denial policies for other unsafe topics 1.
评论
还没有评论,欢迎留下第一条。