Two new policy objectives
Anthropic added two policy goals to its Frontier Safety Roadmap on October 19, 2026. One targets a proposal by December 15 on government support for embedded independent evaluators, including funding, deeper access to models and internal information, and protection of independence.
The second addresses the pace of AI development. Anthropic argues that automated AI research could outrun safety measures and governance. By January 15, 2027, it intends to publish policy options or explain why it decided not to. These are company-defined future objectives, not government-adopted restrictions, legal duties or a development pause.
Ask what evaluators can actually inspect
For buyers, the announcement highlights questions behind an “externally evaluated” label: what access evaluators receive, who funds them, how independence is protected, and which findings and follow-up actions become public. The content and effectiveness of unpublished proposals cannot yet be assessed. Because the roadmap explicitly allows changes, its target dates should be checked against subsequent publications and implementation evidence. A stated ambition is different from a control already operating in a deployed service.