Anthropic's Responsible Scaling Policy (RSP) version 2.2 took effect May 14, 2025. The RSP is Anthropic's public commitment not to train or deploy models capable of causing catastrophic harm unless safety and security measures keep risks below acceptable levels. It was superseded by RSP v3.1 (Anthropic's Responsible Scaling Policy (Version 3.1)).
Core framework
The RSP is built on AI Safety Level (ASL) Standards — technical and operational measures for safely training and deploying frontier AI models. These fall into two categories: Deployment Standards, which set requirements for how models are deployed and used, and Security Standards, which set requirements for protecting model weights from theft. As model capabilities increase, successively higher ASL Standards apply. All v2.2-era models were required to meet the ASL-2 Deployment and Security Standards.
Movement from ASL-2 to ASL-3 is triggered by capability thresholds, which RSP v2.2 defines in two domains: CBRN (Chemical, Biological, Radiological, Nuclear), covering assistance with weapons of mass destruction, and Autonomous AI R&D, covering the capability to autonomously conduct AI research and development. A model whose capabilities are sufficiently far below the thresholds continues under ASL-2; a model that approaches or crosses a threshold requires ASL-3 before deployment.
The assessment process runs in stages. A preliminary assessment determines whether comprehensive testing is needed. Comprehensive testing then evaluates whether the model is sufficiently below the thresholds. Where the outcome is uncertain, the model is treated as having surpassed the threshold, ASL-3 is implemented, and follow-up is conducted. A safeguards assessment evaluates robustness to misuse (Deployment) and resistance to weight theft (Security).
On governance, the RSP maintains a Responsible Scaling Officer position, provides an anonymous reporting channel for staff to flag potential noncompliance, commits to public release of key evaluation materials with sensitive information removed, and solicits external expert input.
Relation to RSP v3.1
RSP v2.2 was superseded by RSP v3.1 (Anthropic's Responsible Scaling Policy (Version 3.1)), which added ASL-level definitions, expanded the AI R&D threshold specifics, and refined the governance structure. The core ASL framework and capability-threshold logic remained continuous.
Key claims
Anthropic frames risk governance for rapidly evolving AI as something that should be "proportional, iterative, and exportable" (Anthropic, 2025). As of May 2025, all Claude models were deployed under the ASL-2 Standard; ASL-3 deployment applied to later models under RSP v3.1.
Relationships
- superseded-by: Anthropic's Responsible Scaling Policy (Version 3.1) — RSP v3.1 is the current version
- related: AI Safety Cases and Frameworks — RSP is Anthropic's primary safety framework
- related: CBRN Uplift — CBRN capability threshold is the primary trigger
- related: Anthropic — author and subject