September 15, 2026
AI safety gets a calendar: Anthropic sets its first milestone for September 30, 2026
Anthropic has set the first Frontier Safety Roadmap milestone for 30 September 2026. By then, the company plans to draw up a list of components, costs, and timelines for an isolated research environment and build a provable inference prototype that ties model output to a specific set of weights.
In April, Anthropic completed its data-retention principles goal, then in May removed it from the current plan and moved the first “moonshot R&D” phase from 15 May to 30 September. The roadmap now lists the following milestones: 1 October 2026 for a systematic evaluation of Claude against its Constitution, and 1 January 2027 for automated cyberattack investigations. For practitioners, this is a calendar of promised evaluations, not an independent measurement of the model.
By 1 January 2027, Anthropic plans to automate cyberattack investigations with minimal human involvement.