Anthropic added datacenter physical security and platform security to its public Frontier Safety Roadmap in an October 19 update. The company tied the changes to its growing datacenter buildout and deployments on infrastructure it operates more directly.
The dated change log makes it possible to distinguish a new commitment from a goal that was already on the page. It also records completed, changed, and retired work.
The October update also added goals concerning embedded third-party evaluators and a policy framework for the pace of AI development. Those are future policy deliverables, rather than proof that an independent evaluator has endorsed every deployed model.
The added policy deadlines are specific: a proposal on embedded third-party evaluators is due December 15, 2026, and a framework addressing the pace of AI development is due January 15, 2027.
Anthropic's separate October 15 evaluator announcement describes the access already being offered to METR and the publication and redaction terms attached to it. The operating arrangement establishes terms for an existing relationship. The future proposal is a commitment to produce a broader policy document by December 15.
A roadmap records commitments, rather than completed safeguards. Its original wording, deadlines and linked evidence allow comparison with later reports. Because the live page changes, a dated version is needed to reconstruct what the company promised at a particular time.
