Anthropic Responsible Scaling Policy

Anthropic has published its Responsible Scaling Policy (RSP), a framework of technical and organizational protocols designed to manage the catastrophic risks associated with the development of increasingly capable AI systems. The policy establishes a structured approach to scaling AI capabilities only when corresponding safety and security measures are proven effective.

The AI Safety Level (ASL) Framework

Anthropic's RSP centers on the AI Safety Level (ASL) system, which is modeled after the US government's biosafety level (BSL) standards. This framework requires that safety, security, and operational standards scale in proportion to a model's potential for catastrophic risk.

ASL Definitions and Criteria

  • ASL-1: Systems that pose no meaningful catastrophic risk, such as a 2018 LLM or an AI system dedicated to playing chess.
  • ASL-2: Systems showing early signs of dangerous capabilities (e.g., providing instructions on building bioweapons) but where the information is not yet useful due to insufficient reliability or because the information is already available via search engines. Current LLMs, including Claude, are categorized as ASL-2.
  • ASL-3: Systems that substantially increase the risk of catastrophic misuse compared to non-AI baselines (such as textbooks or search engines) or systems that demonstrate low-level autonomous capabilities.
  • ASL-4 and ASL-5+: These levels are not yet defined due to their distance from current system capabilities, but they are expected to involve qualitative escalations in autonomy and catastrophic misuse potential.

Safety Measures and Implementation

Safety requirements increase in rigor as a model moves up the ASL scale. ASL-2 measures align with Anthropic's previous White House commitments. ASL-3 requires stricter standards, including enhanced security requirements and a commitment not to deploy models if they exhibit meaningful catastrophic misuse risk during adversarial testing by world-class red-teamers.

For ASL-4, Anthropic commits to defining the measures before reaching ASL-3. These measures may involve solving currently unsolved research problems, such as using interpretability methods to mechanistically demonstrate that a model is unlikely to engage in catastrophic behaviors.

Operational and Business Impact

The RSP is designed to balance risk mitigation with the pursuit of beneficial AI applications. It creates a mechanism where training for more powerful models may be temporarily paused if AI scaling outstrips the ability to comply with safety procedures. This structure is intended to incentivize the resolution of safety issues to unlock further scaling.

From a business perspective, the RSP does not alter the current use or availability of Claude. Anthropic compares the policy to pre-market testing in the aviation or automotive industries, where safety is rigorously demonstrated before a product is released to the market.

Governance and Iteration

Anthropic's RSP has been formally approved by its board, and any changes to the policy must be board-approved following consultations with the Long Term Benefit Trust. The policy is an early iteration and is expected to evolve through rapid course correction as the field of AI progresses.

Anthropic acknowledged the influence of ARC Evals, specifically regarding their expertise in AI risk assessment and the development of the broader ARC Responsible Scaling Policy framework.

Sources

Related

  • Dispatch
    Anthropic Responsible Scaling Policy (RSP) Framework

    Anthropic has introduced a Responsible Scaling Policy (RSP) that uses AI Safety Levels (ASL) to trigger specific safety safeguards as AI models acquire dangerous capabilities.

  • Dispatch
    Anthropic Responsible Scaling Policy Update

    Anthropic has updated its Responsible Scaling Policy (RSP) to introduce a more flexible risk governance framework with specific capability thresholds and proportional AI Safety Level (ASL) standards.

  • Dispatch
    Anthropic Responsible Scaling Policy Reflections

    Anthropic provides an operational update on its Responsible Scaling Policy (RSP), detailing the implementation of AI Safety Levels (ASL) and the frameworks used to mitigate catastrophic risks in frontier models.

  • Dispatch
    Anthropic Responsible Scaling Policy Version 3.0

    Anthropic has released Version 3.0 of its Responsible Scaling Policy (RSP), restructuring its framework to separate unilateral company commitments from industry-wide safety recommendations while introducing a Frontier Safety Roadmap and periodic Risk Reports.

  • Dispatch
    Anthropic Activating AI Safety Level 3 (ASL-3) Protections

    Anthropic has activated AI Safety Level 3 (ASL-3) protections for Claude Opus 4 as a precautionary measure to mitigate risks associated with CBRN weapons development and model weight theft.