Anthropic's responsible scaling policy
On 19 September 2023 Anthropic published its Responsible Scaling Policy (RSP): AI Safety Levels ASL-1 to ASL-3, modelled on biosafety levels, with higher levels requiring stricter demonstrations of safety. The company placed current models, Claude included, at ASL-2; ASL-4 and above were not yet defined.
Why it matters
A laboratory tied further scaling to measured dangerous capabilities: the policy implicitly requires a temporary pause in training more powerful models if scaling outstrips the ability to put the required measures in place. Within months OpenAI published a similar framework, and in 2026 this policy was rewritten as version 3.0.
ASL-2 means early signs of dangerous capabilities, such as instructions on bioweapons that are not yet useful for lack of reliability; ASL-3 a substantial increase in the risk of catastrophic misuse over a search engine or textbook, or low-level autonomy. At ASL-3 the company commits not to deploy a model if adversarial testing by world-class red-teamers shows meaningful misuse risk, and to write the ASL-4 measures before reaching ASL-3. The board approved the policy, and changes are approved by the board after consultation with the Long Term Benefit Trust. ARC Evals' framework is named as the inspiration. Application: in March 2024 Anthropic determined Claude 3 to be ASL-2 'per our Responsible Scaling Policy', which is the same maker; no independent check of compliance was found. The full policy document was not opened; the record stands on the announcement.