Our Responsible Scaling Policy: A Commitment to Safety-Capability Balance
We're publishing our Responsible Scaling Policy, a framework that ties capability increases to corresponding safety advances. No frontier model ships until it passes defined safety thresholds.
Today we’re publishing our Responsible Scaling Policy (RSP), a framework that governs how we develop and deploy increasingly capable models.
Core Principle
Capabilities and safety must advance together. We will not release models that achieve significant capability thresholds without corresponding advances in safety evaluations, interpretability, and alignment techniques.
Safety Levels
We define four safety levels (ASL-1 through ASL-4) tied to specific capability thresholds. Each level has required mitigations before deployment is permitted.
Enforcement Mechanism
The RSP is enforced by an independent safety committee with binding authority to delay or cancel model releases. This committee includes external researchers not employed by AXIOM.
Our Commitment
This policy represents our commitment to the belief that the development of powerful AI can and must be done responsibly. We publish this document not because it is legally required, but because we believe in transparency as a core value.