Anthropic says it met September 30 deadline for Frontier Safety Roadmap Phase 1 goal

Anthropic's Frontier Safety Roadmap page, in an entry dated October 19, 2026, says the company broadly achieved the goals set for Phase 1 of its moonshot R&D security projects, whose deadline was September 30, 2026. The target had originally been May 15, 2026 and was moved in May so that resources could go to a separate goal on strengthening security across the board. The page presents the roadmap as a public statement of priorities in security, safeguards, alignment and policy.

On the first project, Anthropic says it built a prototype showing how key workflows and infrastructure would look under extreme security practices and examined the systems needing the longest lead time to support air-gapped operation. It admits most effort went into prototype environments for measuring productivity cost, not physical security controls, and says it plans to run inference in isolated environments soon. On the second, it built a prototype that checks production inference outputs against a specific set of model weights, and signing of internal token traffic has already caught one rare, benign glitch. The company now calls this work inference verification instead of provable inference.

The same update records other changes. An automated red-teaming system finished a goal early, matching or beating the best human bounty submission in all 24 question-classifier pairs and winning 23, within 24 to 48 hours. The goal on upholding Claude's Constitution was retired in its current form. Two policy goals were added: one proposal covering evaluators embedded from outside, due December 15, 2026, and a framework addressing how fast AI is developed, due January 15, 2027.