ASL-3 Protective Measures Activated

Alongside the launch of Claude Opus 4, Anthropic has activated the AI Safety Level 3 (ASL-3) deployment and security standards specified in its Responsible Scaling Policy (RSP).

This measure includes enhanced security to prevent theft of the model's weights, as well as deployment measures to prevent the model from being misused for the development or acquisition of chemical, biological, radiological, and nuclear (CBRN) weapons.

Anthropic has defined this application as a precautionary and provisional measure. It has not yet been confirmed whether Claude Opus 4 has definitively crossed the 'Capability Threshold' that would make ASL-3 application necessary, but because the model's CBRN-related knowledge and capabilities continue to improve, Anthropic judged that the possibility of risk cannot be completely ruled out.

The key components of the ASL-3 standard are as follows:

• Security Controls: Enhanced internal security measures to protect model weights from attacks by sophisticated non-state actors.
• Deployment Measures: Precise deployment controls designed to limit CBRN-related risks.

Through ongoing evaluation going forward, if it is determined that Claude Opus 4 has not crossed the relevant threshold, Anthropic may adjust or remove the ASL-3 protective measures.