TechnologyTracking
Can Anthropic's independent evaluation pledge deliver?
After granting METR employee-level access and waiving publication veto, can its independent assessments truly reveal frontier AI safety risks without being materially constrained by commercial interests or confidentiality limits?
Watching: METR 尚未发布任何基于该访问权限的评估报告，其实际独立性和披露程度未知；Anthropic 的删减权范围（如模型架构细节）可能涵盖关键安全信息，但具体执行情况不明。
Current read
Anthropic has granted METR employee-level access, but no report has been published; independence remains uncertain due to confidentiality redactions and commercial interests.
According to Anthropic's October 15, 2026 announcement, the company has provided six METR employees with company hardware, office space, and internal tools and system access comparable to its own risk assessment team; METR will publish its risk assessment results. Access excludes a few defined categories such as customer data, attorney-client privileged information, and model architecture, and Anthropic may request redactions within these limits. This mechanism aims to strengthen independent evaluation, but confidentiality redactions may limit disclosure of critical safety information, and commercial interests could influence the process. As of now, METR has not published any report based on this access, so actual independence and disclosure levels remain unknown.
How we read it
Anthropic has granted METR employee-level access including hardware, office space, and internal tools, but no report has been published yet.
Access excludes customer data, attorney-client privilege, and model architecture, with Anthropic able to request redactions in these areas, potentially affecting assessment completeness.
METR will publish its risk assessment results, but specific content and redaction implementation are unknown, leaving actual independence to be observed.
Earlier updates
- — Public tracking started; no independent probability estimate yet.; Starting public context: Anthropic has granted six METR researchers employee-level access to independently assess its AI models, safety practices, and risk management, providing company laptops, office space, and internal tools comparable to its own risk assessment teams. METR can publish its assessments without Anthropic's approval, though Anthropic may require redactions in narrow predefined categories such as customer data, attorney-client privileged information, and model architecture details. METR sits at the top of a three-tier evaluator system Anthropic is building, with other evaluators like SecureBio at the lowest tier requiring Anthropic's approval to publish.
- — Anthropic has granted METR employee-level access, but no report has been published; independence remains uncertain due to confidentiality redactions and commercial interests.; According to Anthropic's October 15, 2026 announcement, the company has provided six METR employees with company hardware, office space, and internal tools and system access comparable to its own risk assessment team; METR will publish its risk assessment results. Access excludes a few defined categories such as customer data, attorney-client privileged information, and model architecture, and Anthropic may request redactions within these limits. This mechanism aims to strengthen independent evaluation, but confidentiality redactions may limit disclosure of critical safety information, and commercial interests could influence the process. As of now, METR has not published any report based on this access, so actual independence and disclosure levels remain unknown.