USGuidelineProposedModel providersDeployers

Allowing third-party evaluators permanent, employee-level system access

允许第三方评估者获得永久、员工级别的系统访问权限

Anthropic, OpenAI·Published 2026-09-13·Checked by 管理员 on 2026-09-15

Anthropic and OpenAI separately announced that they will allow third-party evaluators permanent, employee-level system access to audit safety measures, report incidents, and assess model alignment during training, and advocate for slowing down the development of frontier models.

Source ↗Remind me US · All

What it means for you

• Obligations: AI companies must open their systems to third-party evaluators for safety audits and model alignment assessments, increasing compliance costs and operational complexity.
• Opportunities: Proactively accepting third-party evaluations can enhance transparency, build trust with the public and regulators, and align with industry self-regulation trends.
• Deadlines: Companies need to assess and implement third-party evaluation mechanisms, potentially impacting future model development plans.

What to do

ComplianceAssess and develop a specific plan to allow third-party evaluators to gain employee-level system access, including technical interfaces, security protocols, and processes.

ComplianceCommunicate with potential third-party evaluation institutions to clarify the scope, frequency, and standards of the evaluation, ensuring they have the capability and meet the requirements.

ProductReview and update internal model development processes to ensure they align with the industry initiative to 'slow down the development of frontier models' and be prepared for external evaluation.