Dario Amodei, CEO of Anthropic, has proposed a new model for AI safety oversight. In an essay published on September 12, 2026, titled We Must Pace the Frontier, he outlined a plan to integrate independent third-party evaluators directly into the company's development process.
Under this framework, external reviewers would receive the same access as full-time employees. This includes dedicated desks in Anthropic offices, company access badges, and corporate laptops, allowing them to monitor AI development environments continuously rather than through periodic audits.
A key component of the proposal is the autonomy granted to these evaluators. They would be permitted to publish their findings with minimal redaction, and Anthropic would not exercise editorial control over the results of their assessments.
Amodei identified METR, an organization specializing in AI safety evaluation, as a potential partner for this initiative. The model aims to replace traditional, episodic audits with continuous observation to identify potential issues during the development phase.
This initiative marks a transition from advocating for independent testing to implementing operational commitments. The industry will monitor how these evaluators handle proprietary training methods and whether the publication rights remain meaningful in practice.
