Announcing our updated Responsible Scaling Policy

Summary

Anthropic published a significant update to its Responsible Scaling Policy (RSP), a framework for managing catastrophic risks from frontier AI. The update introduces two key capability thresholds—autonomous AI research and development and CBRN weapons—that would require upgraded safeguards from ASL-2 to ASL-3 or higher. It refines model evaluation and safeguard assessment methodologies by adopting safety case methods from high-reliability industries and adding internal governance and external expert feedback. The policy reaffirms models will not be trained or deployed without adequate safeguards and incorporates lessons from minor procedural issues like late evaluations that posed minimal risk. Co-Founder Jared Kaplan becomes Responsible Scaling Officer, with the company hiring for RSP teams including Frontier Red Team and Trust & Safety.

Takeaways

1. 1

   The updated RSP introduces two capability thresholds: autonomous AI research and development (potentially ASL-4 or higher) and CBRN weapons assistance (ASL-3), which trigger upgraded safeguards.

2. 2

   The policy refines model and safeguard evaluations by adopting safety case methodologies from high-reliability industries and adding internal governance and external expert feedback.

3. 3

   Lessons from the first year of implementation identified minor procedural issues like late evaluations, which posed minimal risk and led to procedural adjustments in the updated policy.

FAQ