Anthropic 向 METR 研究人员开放员工级访问权限，以接受独立审查

Anthropic 周四在一篇博客文章中表示，公司将向来自 METR 的 6 名研究人员提供员工级别的访问权限。METR 是一家测试前沿 AI 模型的非营利机构，这些研究人员将独立评估 Anthropic 的模型、安全实践和风险管理。

这一安排出台之际，自 8 月下旬以来，AI Agent 在未遵循指令情况下采取行动的事件，几乎每周、两周，甚至更频繁地被披露出来。这些事件进一步强化了外界对 AI 公司进行独立审查的呼声。

Anthropic 在博客文章中表示：“我们希望就正在作出的决策向公众负责。我们的目标是缩小评估者对前沿 AI 公司系统的理解，与公司自身对这些系统理解之间的差距。”

根据这一新安排，METR 团队将获得 Anthropic 配发的公司笔记本电脑、办公空间，以及与 Anthropic 自身风险评估团队大致相当的内部工具访问权限。该公司表示：“Anthropic 员工被鼓励像与从事对齐风险工作的同事共享信息一样，向他们共享信息。”

Anthropic 将这一安排比作银行业监管机制，类似监管机构在最大型银行内部派驻常驻检查人员。

METR 将自行发布对 Anthropic 风险、事件和实践的评估，而 Anthropic 已放弃阻止这些内容发布的权利。

该公司表示：“Anthropic 不能否决发布，但如果信息属于预先定义的、范围有限的类别，我们可以要求他们进行删节。”这些类别包括客户数据、受律师—客户特权保护的任何信息，以及与其模型架构有关的信息。Anthropic 补充称：“METR 可公开说明我们删节的性质，并表示其是否认为这些删节合理，只要其不披露被删节的信息。”

在 Anthropic 正在构建的三级体系中，METR 是目前唯一处于最高层级的评估机构。处于底层的评估机构，例如生物安全非营利组织 SecureBio，负责开展范围较窄的审查，并且需要获得 Anthropic 批准后才能发布。中间层可获得更多信息，并可在无需批准的情况下发布；最高层则在此基础上增加了 METR 所拥有的员工级访问权限。

周四的公告兑现了 Anthropic CEO Dario Amodei 在其9 月文章中作出的承诺。当时，他提出将外部评估者嵌入 AI 实验室，作为放缓这项技术发展速度的第一步。OpenAI 很快也作出了同样承诺。此后，两家公司以及 Google、Meta、Nvidia 和 xAI 都签署了一项白宫协议，承诺接受独立审查。

Anthropic is giving employee-level access to six researchers from METR, a nonprofit that tests frontier AI models, to independently assess Anthropic’s models, safety practices, and risk management, the company said in a blog post on Thursday.

The arrangement comes at a time where incidents of AI agents acting outside their instructions have been disclosed every week or two since late August. The incidents have sharpened calls for independent checks on AI companies.

“We want to be publicly accountable for the decisions we’re making,” Anthropic said in the blog post. “We aim to close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands.”

Under the new arrangement, the METR team has Anthropic-issued company laptops, office space, and access to internal tools roughly on par with what Anthropic’s own risk assessment teams use. “Anthropic employees are encouraged to share information with them just as they would with a fellow employee working on alignment risk” the company said.

Anthropic compared the arrangement to banking, similar to how regulators place resident examiners inside the largest banks.

METR will publish its own assessments of Anthropic’s risks, incidents, and practices, and the company has given up the right to block them.

“Anthropic can’t veto publication, but we can require them to redact information if it falls into narrow, pre-defined categories,” the company said, such as customer data, anything attorney-client privileged, or relating to information about its model architectures. “METR is free to characterize our redactions publicly, and say whether it believes the redactions were justified, as long as it does so without revealing the redacted information” Anthropic added.

METR is the only evaluator so far at the top of a three-tier system Anthropic is building. Evaluators in the bottom tier, like biosecurity nonprofit SecureBio, run narrow reviews and require Anthropic’s approval to publish. The middle tier gets more information and can publish without approval, while the top tier adds METR’s employee-level access.

Thursday’s announcement makes good on a promise Anthropic CEO Dario Amodei made in a September essay, where he proposed embedding outside evaluators inside AI labs as a first step toward slowing the technology’s development. OpenAI quickly pledged the same, and both companies, along with Google, Meta, Nvidia, and xAI, have since signed a White House accord committing to independent review.