Skip to content
worth noting Security

Anthropic has released version 3.0 of its Responsible Scaling Policy safety framework

only one source so far

Anthropic has released the third version of its voluntary Responsible Scaling Policy framework for managing catastrophic AI risks. It builds on more than two years of experience with the original 2023 policy and, according to the company, is meant to increase transparency and accountability in decision-making.

Anthropic has released the third version of its Responsible Scaling Policy (RSP) – a voluntary framework used to manage catastrophic safety and misuse risks associated with its frontier models. The update builds on more than two years of experience with the original policy from September 2023 and, according to the company, is meant to reinforce what has worked, fix shortcomings, and introduce new measures for greater transparency and accountability in decision-making.

The original RSP was built on conditional ("if-then") commitments tied to AI Safety Level (ASL) tiers: if a model exceeded a certain capability (for example in the biological domain), the company committed to implementing stricter safety measures corresponding to a higher ASL level. The company describes four mechanisms through which it aimed to influence the entire industry: an internal enforcement tool for its own organization, an effort to create a "race to the top" among companies, building consensus on risks at key capability thresholds, and preparing for future coordination with governments on risks that no single company can handle alone.

According to the company's own assessment, part of this plan has been realized: it developed more sophisticated classifiers for blocking dangerous content, activated the ASL-3 standard for relevant models in May 2025, and similar frameworks were adopted by OpenAI and Google DeepMind within a few months. According to the company, principles from the RSP have also influenced emerging regulation – California's SB 53 law, New York's RAISE Act, and codes of practice under the EU AI Act now similarly require developers of frontier models to publish frameworks for managing catastrophic risks.

The company's earlier reflection on implementing the policy also mentioned the concept of "Red Line Capabilities" – future model capabilities that, under today's safety practices, would pose too high a risk, and testing models for their presence through so-called Frontier Risk Evaluations. The specific changes to commitments that version 3.0 brings compared to the previous version are not apparent from the available excerpt. For details, see the source article.

What changed

Why it matters

The framework itself is voluntary and applies primarily to Anthropic's own models, but it influences the broader regulatory landscape – according to the company, similar obligations to publish risk frameworks have already found their way into California's SB 53, New York's RAISE Act, and EU AI Act codes of practice. For companies dealing with compliance when deploying AI or assessing vendor risk, this is a signal of where regulation of frontier models is heading.

What was added since the original report

Verified updates

  1. New verified information

    Anthropic introduced the concept of Red Line Capabilities – future capabilities posing too high a risk under current safety practices; frontier models are tested for the presence of Red Line Capabilities, and failure to rule them out leads to the assumption that they are present; implementation of the RSP identified important questions, projects, and dependencies within the safety framework; Anthropic integrates practices from nuclear safety, biosecurity, autonomous vehicles, aviation, and cybersecurity; Anthropic is working on an updated version of the Responsible Scaling Policy.

    • Anthropic introduced the concept of Red Line Capabilities – future capabilities posing too high a risk under current safety practices
    • Frontier models are tested for the presence of Red Line Capabilities; failure to rule them out leads to the assumption that they are present
    • Implementation of the RSP identified important questions, projects, and dependencies within the safety framework
    • Anthropic integrates practices from nuclear safety, biosecurity, autonomous vehicles, aviation, and cybersecurity
    • Anthropic is working on an updated version of the Responsible Scaling Policy

Relevant practical impact

What this means

01

For a business

Companies deploying Anthropic's models or tracking AI regulation can use RSP 3.0 as a reference when assessing vendor risk and tracking emerging legal requirements (SB 53, RAISE Act, EU AI Act), which already require similar frameworks.

Risks and compliance
What to decide Study the full text of RSP 3.0 if a company is assessing risks associated with using Claude models or tracking related regulatory requirements.
More business impacts →
Anthropic AI safety frontier models Red Line Capabilities Responsible Scaling Policy

Check the original

Event sources

only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.

2
Anthropic News primary source · first detected Reflections on our Responsible Scaling Policy Anthropic News primary source Responsible Scaling Policy Version 3.0