Responsible Scaling Policy

BLUF

The Responsible Scaling Policy (RSP) is Anthropic’s internal governance framework for managing the dual-use risks of increasingly capable frontier AI models. Introduced in 2023 and updated in 2024, it establishes a tiered “AI Safety Level” (ASL) classification system — analogous to biosafety levels (BSL-1 through BSL-4) — that determines what deployment, access, and research restrictions apply to models at each capability tier. The RSP is the primary public reference for Anthropic’s self-regulatory commitments; it governs (in principle) whether and how models like Claude Mythos may be deployed, under what access controls, and with what usage restrictions.


ASL Classification System

LevelRisk ThresholdImplication
ASL-1No meaningful uplift beyond publicly available toolsNo specific deployment restrictions
ASL-2Meaningful uplift to CBRN or critical-infrastructure attackCurrent deployment standard; enhanced monitoring required
ASL-3Substantial uplift to mass-casualty weapon development or autonomous cyberattackRestricted deployment; government notification; mandatory safety measures
ASL-4Catastrophic potential (autonomous replication, major weapons development)No deployment without extraordinary safety measures

As of 2024, Anthropic has classified Claude 3 and Claude 3.5 models as ASL-2. No public ASL classification for Claude Mythos has been released.


Analytical Significance

The RSP represents an attempt to self-regulate frontier AI development in the absence of binding government regulation. Its limitations are analytically significant:

  • Self-reporting: Anthropic determines its own ASL classification without independent audit or mandatory government disclosure
  • Enforcement gap: No legal mechanism compels compliance; the RSP is a policy commitment, not a legal obligation
  • Mythos gap: No published RSP evaluation for Claude Mythos exists — the most capable Anthropic model as of April 2026 lacks a public capability-risk assessment, even as NSA operationalizes it for vulnerability research
  • Comparison: OpenAI’s Preparedness Framework is structurally similar; both are voluntary self-governance instruments; neither has binding external enforcement

Key Connections

  • Anthropic — policy author and sole enforcer
  • Claude Mythos — the model for which no public RSP evaluation exists
  • Constitutional AI — Anthropic’s alignment methodology; the technical basis for RSP safety claims
  • Dual-Use Technology — RSP is specifically designed to manage dual-use AI capability risks

Sources

  • Anthropic RSP documentation (anthropic.com/responsible-scaling-policy) — [High confidence — primary]
  • UK AI Safety Institute evaluations (2023–2024) — [High confidence]
  • Center for AI Safety (CAIS): RSP analysis — [Medium confidence]