UK AI Safety Summit

AI technology is advancing very rapidly, with computing resources growing 8x every year. This rapid progress makes it difficult to predict when AI will acquire dangerous capabilities, such as the ability to manufacture biological weapons.

To address this uncertainty, Anthropic has introduced a Responsible Scaling Policy (RSP). This policy operates around two core pillars: the AI Safety Levels (ASL) framework and regular risk testing.

The main stages of ASL (AI Safety Levels) are as follows:

• ASL-1: A stage for special-purpose AI with almost no risk.
• ASL-2: The current stage, applying best security practices such as model card creation and external red-teaming activities.
• ASL-3: A stage where catastrophic misuse is possible in the CBRN (Chemical, Biological, Radiological, Nuclear) domain. Requires strong security along with technical breakthroughs to prevent the generation of dangerous information.
• ASL-4: A stage where AI could escape human control or become a serious global security threat. Requires a precise understanding of the model's internal workings as a prerequisite.