TechnologyTracking
Can Anthropic's internal access for METR set a new standard for AI safety evaluation?
Can Anthropic's granting of internal-level access to six METR safety evaluators effectively constrain frontier AI safety practices and become an industry standard?
Watching: METR评估的实际效果未知，Anthropic的删节权可能限制透明度，其他公司是否跟进尚不确定。
Current read
Anthropic granted six METR evaluators employee-level access; METR can publish independently but with redaction limits. Whether this becomes an industry standard remains uncertain.
Reports from October 16, 2026 state that Anthropic has granted six METR evaluators company hardware, office space, and access roughly on par with its internal risk teams. METR can publish findings independently; Anthropic cannot veto but can request redactions for categories like customer data, legal privilege, and model architecture, with METR able to publicly challenge redactions. Anthropic compares this to bank regulators' resident examiners and uses a three-tier evaluation system with METR at the top. The reports do not specify the content or release date of METR's first assessment, nor whether other frontier AI companies will adopt similar arrangements. The model's effectiveness and industry adoption remain uncertain.
How we read it
Anthropic granted six METR evaluators employee-level access; METR can publish independently, but Anthropic retains redaction rights for specific categories, which METR can publicly challenge.
Anthropic likens the arrangement to bank regulators' resident examiners and uses a three-tier system with METR at the top; reports do not provide METR's first report content or other companies' follow-up.
Earlier updates
- — Public tracking started; no independent probability estimate yet.; Starting public context: METR will publish its own verdicts on the risks of Anthropic's models and practices, which Anthropic cannot block and may only redact in a few predefined categories such as customer data, a decision METR can publicly challenge. Other evaluators such as SecureBio begin at the lowest of three new access tiers, where Anthropic still controls what they see and must approve what they publish.
- — Anthropic granted six METR evaluators employee-level access; METR can publish independently but with redaction limits. Whether this becomes an industry standard remains uncertain.; Reports from October 16, 2026 state that Anthropic has granted six METR evaluators company hardware, office space, and access roughly on par with its internal risk teams. METR can publish findings independently; Anthropic cannot veto but can request redactions for categories like customer data, legal privilege, and model architecture, with METR able to publicly challenge redactions. Anthropic compares this to bank regulators' resident examiners and uses a three-tier evaluation system with METR at the top. The reports do not specify the content or release date of METR's first assessment, nor whether other frontier AI companies will adopt similar arrangements. The model's effectiveness and industry adoption remain uncertain.