A new AI safety playbook

Anthropic published major updates to its Responsible Scaling Policy, the framework it uses to prevent catastrophic risks from advanced AI systems.

Core commitment remains unchanged

The company won't train or deploy models unless adequate safeguards are implemented. Protection scales with potential risks.

AI Safety Levels explained

ASL Standards are graduated safety measures that get stricter as AI capabilities increase. They start at ASL-1 for basic bots and progress upward.

Current models at baseline

All Anthropic models currently operate under ASL-2 Standards, which reflect current industry best practices.

Threshold one: AI research

If a model can independently conduct complex AI research tasks, it requires ASL-4 or higher security standards to prevent development from outpacing safety.

“If a model can independently conduct complex AI research tasks typically requiring human expertise, potentially significantly accelerating AI development in an unpredictable way”

Threshold two: weapons assistance

If a model can help someone with basic technical knowledge create CBRN weapons, it requires enhanced ASL-3 security and deployment controls.

What ASL-3 safeguards include

Enhanced security with internal access controls and model weight protection. Multi-layered deployment approach with real-time monitoring and rapid response protocols.

Routine capability assessments

Models are regularly evaluated against capability thresholds to determine if current safeguards remain appropriate. Summaries are publicly available.

Learning from compliance gaps

First-year review found minor procedural issues like evaluations completed three days late. All instances posed minimal risk but revealed need for more flexibility.

External input and stress-testing

The updated policy includes internal stress-testing, existing safety reporting processes, and solicits feedback from external experts on assessment methods.

New leadership for scaling

Jared Kaplan becomes Responsible Scaling Officer, succeeding Sam McCandlish who led initial implementation. Anthropic is hiring a Head of Responsible Scaling.