Anthropic updates its Responsible Scaling Policy

Background

The first version of the Responsible Scaling Policy (RSP) had set AI Safety Levels (ASL) and rules for when stronger safeguards were needed. A year later Anthropic revised it.^[1]

What happened

On 15 October 2024 Anthropic published a major update to the RSP, version 2.0.^[1][2]

Key details

The update defined capability thresholds for two areas: autonomous AI research and development, and CBRN weapons assistance. Reaching either threshold would trigger stronger safeguards.^[1]

Reaching the CBRN threshold requires ASL-3 standards. Autonomous AI research and development could require ASL-4 or higher security.^[1]

At the time, all Anthropic models operated under ASL-2.^[1]

Anthropic disclosed small shortfalls in its first year, such as evaluations completed three days late.^[1]

Jared Kaplan became Responsible Scaling Officer, succeeding Sam McCandlish, who stayed on as chief technology officer.^[1]

A capability threshold is a level at which a model's abilities are judged strong enough to need extra safeguards. Setting thresholds in advance means that release decisions follow a stated rule rather than being made case by case.

Impact

The update made the policy's triggers more specific and named the people responsible for it. Later versions followed: version 2.1 on 31 March 2025 and version 2.2 on 14 May 2025.^[2]