Anthropic 向非营利机构 METR 开放内部权限，审计其 AI 模型与安全实践
Anthropic gives METR inside access to audit its AI models and safety practices
- 1Anthropic 给予非营利机构 METR 六人团队内部权限，审查其 AI 模型与安全实践
- 2METR 员工获公司硬件、办公空间和内部工具，官方称权限接近其内部风险团队
- 3客户数据、律师保密信息和模型架构细节不开放，METR 将独立发布审查结果
- 4此次开放超过 SecureBio、Accenture、Faculty 此前获得的权限，并紧随白宫超级智能协议签署
Anthropic 将外部安全审计从象征性访问升级为接近内部团队的实际权限，为第三方模型审计立下更深的开放尺度。
- 影响：采购前沿模型的企业可能多出一个比厂商自述更可靠的独立安全信息渠道；不同厂商接受外部审查的深度差异，可作为供应商安全尽调的对比项；模型架构等商业敏感信息仍是透明度边界。
- 参考：可考虑把第三方审计安排纳入模型选型的尽调清单，比较各厂商外部审查深度与发布限制；已采用 Anthropic 模型的团队值得留意 METR 后续独立报告。
- 依据：开放范围与删减机制为 Anthropic 官方口径，经 The Decoder 转述；METR 审查结果尚未发布，暂无独立验证。
Anthropic has given a team from the nonprofit METR inside access to review its AI models and safety practices, the company announced. Six METR staffers received company hardware, office space, and access to internal tools, which Anthropic says puts them roughly on par with its own risk teams.
Customer data, information covered by attorney-client privilege, and details about the model architecture remain off-limits. METR will publish its findings independently. Anthropic says it can ask for redactions in certain predefined categories, but METR is free to say publicly whether it thinks those redactions were justified.
Anthropic argues that not all safety-relevant information can be made public, either because it could help attackers or because it's commercially sensitive. The METR deal goes further than what the company has offered other outside reviewers. SecureBio, Accenture, and Faculty have so far received less access and tighter limits on what they can publish, and Anthropic has started early talks with Hugging Face. The move follows Anthropic's signing of the White House Accord on Super Intelligence, which includes a commitment to working with outside reviewers.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
来源：The Decoder：AI News · the-decoder.com