Skip to main content
Zubnet AI › Learn › Wiki › Responsible Scaling Policy
Safety

Responsible Scaling Policy

Also known as: RSP, Frontier Safety Framework, Preparedness Framework
A public commitment by an AI developer to test its models for dangerous capabilities and to have specific safeguards in place before developing or releasing models that cross defined risk thresholds.

Why it matters

These frameworks are the main way frontier AI companies currently govern their own most serious risks. They determine which tests a model goes through before release, what security surrounds its weights and, in principle, when a company would hold a model back.

They've also become the template for regulation. At the AI Seoul Summit in May 2024, sixteen companies pledged to publish such frameworks. Since then, laws and codes have begun turning voluntary practice into obligation.

Deep Dive

A responsible scaling policy is built on if-then commitments: if our model reaches capability X, then we will have safeguards Y in place. Anthropic's original policy also promised to pause until those safeguards were ready, a commitment it dropped in RSP v3.0 (2026). The idea is to tie safety measures to what models can actually do, instead of arguing in advance about when dangerous AI will arrive.

"Responsible Scaling Policy" (RSP) is Anthropic's name for its framework, first published in September 2023 and based on a concept developed with the evaluation nonprofit METR. Other labs use different names for the same basic idea. OpenAI has its Preparedness Framework and Google DeepMind its Frontier Safety Framework, and the generic term is frontier safety framework.

The Common Structure

  • Risk domains: the frameworks focus on severe, large-scale harms. The usual list is help with chemical, biological, radiological and nuclear (CBRN) weapons, advanced cyberattacks, and AI that can autonomously accelerate AI research or evade human control.
  • Capability thresholds: defined levels at which a model would meaningfully increase risk, for example by giving real uplift to someone trying to build a bioweapon.
  • Evaluations: tests run before and during deployment to check whether a model is approaching a threshold, including red teaming and studies with human experts.
  • Safeguards: two kinds. Deployment safeguards limit misuse through classifiers, monitoring and access controls. Security safeguards protect model weights from theft.
  • Governance: who decides, how exceptions are handled, what internal oversight exists and what gets published.

Anthropic's ASL System

Anthropic's RSP defined AI Safety Levels (ASL), loosely modeled on the biosafety levels used for laboratories that handle dangerous pathogens. ASL-2 covers models that don't pose significant catastrophic risk. ASL-3 applies to models that could substantially help people with basic technical backgrounds create CBRN weapons, and it requires stronger security and targeted deployment protections. Anthropic activated ASL-3 protections for the first time in May 2025, with the release of Claude Opus 4. RSP v3.0 (24 February 2026) rewrote the policy: it no longer defines escalating ASL tiers or the pause commitment, but it keeps the ASL-3 safeguards and adds Risk Reports.

Other Frameworks

OpenAI's Preparedness Framework (first published in December 2023, revised in 2025) tracks categories including biological and chemical capability, cybersecurity and AI self-improvement, with "High" and "Critical" thresholds. Google DeepMind's Frontier Safety Framework (May 2024, updated since) defines Critical Capability Levels and has added risks such as harmful manipulation and misalignment. Meta, xAI, Microsoft, Amazon and others have published their own versions.

From Voluntary to Required

  • The EU AI Act's Code of Practice for general-purpose AI (2025) asks providers of the most capable models to maintain a safety and security framework, assess systemic risks and report serious incidents.
  • California's SB 53, signed in September 2025, requires large frontier developers to publish their safety frameworks and report critical safety incidents, and it protects whistleblowers.
  • New York's RAISE Act, signed in December 2025 and amended in March 2026, sets similar requirements that take effect in 2027.

Criticisms

  • Self-policing: companies write the rules, run the tests and judge the results, under strong commercial pressure to ship.
  • Vague thresholds: terms such as "significant uplift" leave wide room for interpretation.
  • Shifting commitments: companies have revised their frameworks in ways that critics say loosened earlier promises, and some frameworks allow safeguards to be adjusted if competitors move ahead without similar protections.
  • Fragile evaluations: tests can underestimate a model because of poor elicitation or, potentially, because the model sandbags.
  • Narrow scope: by design they target catastrophic risks, not everyday harms such as bias, privacy violations or misinformation.

Supporters respond that concrete, public commitments are better than none, that they create internal pressure to invest in safety, and that they gave regulators a tested model to build on.

← All Terms
ESC