Anthropic Drops Responsible Scaling Policy

On February 25, 2026 — one day after Hegseth’s ultimatum — Anthropic published a revised version of its safety framework, formally retiring the Responsible Scaling Policy (RSP) that had been a cornerstone of the company’s public identity since 2023. The new document, titled the “Responsible Deployment Framework,” maintained many of the same evaluation thresholds for catastrophic risk but notably softened language around use-case restrictions and introduced a tiered access model that could theoretically accommodate military applications with different guardrail configurations.

The timing was impossible to separate from the Pentagon crisis. Anthropic leadership insisted the revision had been in development for months and reflected lessons learned from real-world deployments rather than political pressure. Critics were unconvinced. Several prominent AI safety researchers publicly questioned whether the revision represented a genuine evolution in Anthropic’s thinking or a concession dressed up as intellectual progress.

The revised framework did not, however, concede the Pentagon’s core demands. It maintained explicit prohibitions on autonomous weapons systems, mass surveillance without judicial oversight, and AI-driven targeting without human decision-makers in the loop. These were precisely the “red lines” that Dario Amodei would publicly articulate the following day. The RSP revision appeared to be Anthropic’s attempt to demonstrate flexibility on process while holding firm on substance — a distinction the Pentagon was not interested in making.