Anthropic Announces Updated Responsible Scaling Policy (RSP)

Anthropic has updated its Responsible Scaling Policy (RSP), a risk governance framework aimed at mitigating catastrophic risks from frontier AI systems. This update introduces a more flexible and granular approach to risk assessment and management in line with advances in AI technology.

The core principle is Proportional Protection, which strengthens safeguards in proportion to potential risks. To this end, Anthropic uses AI Safety Level Standards (ASL Standards), which apply safety measures in stages according to a model's capabilities.

All models currently follow ASL-2, the industry's highest standard, and reaching certain Capability Thresholds will require higher levels of safety measures, as follows:

• Autonomous AI research and development: When a model can conduct complex AI research without human expertise (requires ASL-4 or higher)
• CBRN (Chemical, Biological, Radiological, Nuclear) weapons: When a model can substantially assist in the creation or deployment of weapons (requires ASL-3)

This policy complements the existing Usage Policy and social impact research, and aims to proactively put safeguards in place to keep pace with the rapid pace of AI advancement.