Anthropic宣布启动独立第三方评估计划，向非营利组织METR的六名研究员授予类似员工级别的访问权限12。这一举措旨在对公司的AI模型、安全实践和风险管理进行独立评估2。获得授权的研究人员将可以访问Anthropic的公司硬件、办公空间以及与公司内部风险评估团队相当的内部工具系统12。
Anthropic在此框架下保有有限的内容审查权限2。METR可自主发布其风险评估分析结果，Anthropic无法否决发布但仅可在客户数据、律师特权信息等狭义范围内要求编辑敏感内容12。这一承诺源自Anthropic CEO Dario Amodei在九月提出的公开许诺2，旨在履行公司对白宫超级智能协议的承诺1。
Anthropic正与多家机构建立合作关系以实现评估目标1。公司已与SecureBio、Accenture、Faculty以及包括Hugging Face在内的其他组织进行评估合作1。该公司建立了三层级的评估者访问框架，其中METR处于最高层级2。有效的评估者需具备评估前沿模型的专业知识、独立性、充分的访问权限和自由发布观察结果的能力1。
Anthropic has launched an independent evaluation program, granting nonprofit organization METR employee-level access to assess its AI models' safety practices and risk management 12. The initiative constitutes part of Anthropic's fulfillment of commitments made through the White House Accord on Super Intelligence, which the company has signed 1. Six METR researchers will receive access to company hardware, office facilities, and internal tools comparable to those available to Anthropic's own risk assessment team 12.
Under the arrangement, METR retains full autonomy to publish its findings, though Anthropic may request edits only for narrowly defined sensitive categories including customer data, attorney-client privileged information, and model architecture details 2. Anthropic has structured a three-tier evaluation framework, with METR positioned at the highest level of access 2. The company is simultaneously developing partnerships with additional evaluation organizations including SecureBio, Accenture, Faculty, and Hugging Face 1.
The framework reflects broader industry movement toward external accountability. Several major AI developers—OpenAI, Google, Meta, Nvidia, and xAI—have similarly committed to independent audits under the White House agreement 2. Anthropic CEO Dario Amodei publicly outlined this evaluation commitment in September 2.
评论
还没有评论，欢迎留下第一条。